diff --git a/README.md b/README.md
index 6c68c27b..4ed97176 100644
--- a/README.md
+++ b/README.md
@@ -69,22 +69,32 @@ const result = await supervise(
 ### Improve an agent
 
 `improve` optimizes one part of an agent and **only ships a change if it beats the current agent on tasks it never practiced on**.
-It accepts prompt, skill document, tool, MCP, hook, subagent, workflow, rollout-policy, whole-profile, and code surfaces through one call.
-Prompt and skill-document optimization have built-in generators; structured surfaces take an explicit generator, and code runs from isolated incumbent and candidate checkouts.
+It accepts prompt, skill document, curated memory, tool, MCP, hook, subagent, workflow, rollout-policy, whole-profile, and code surfaces through one call.
+Prompt, skill-document, memory, and rollout-policy optimization have built-in generators; structured surfaces take an explicit generator, and code runs from isolated incumbent and candidate checkouts.
 
 ```ts
 import { improve } from '@tangle-network/agent-runtime'
 
 const { profile, shipped, lift } = await improve(baseProfile, findings, {
-  surface: 'prompt',        // or skills/tools/mcp/hooks/subagents/workflow/agent-profile/code
+  surface: 'prompt',        // or skills/memory/tools/mcp/hooks/subagents/workflow/agent-profile/rollout-policy/code
   gate: 'holdout',          // certified on a held-back exam, never the practice set
   scenarios, judge, agent,  // how to measure a candidate
 })
 ```
 
+Curated memory is an external lesson document, not a knowledge store. Supply its current text and persist only the promoted winner:
+
+```ts
+await improve(baseProfile, findings, {
+  surface: 'memory',
+  memory: { document: currentLessons, writeBack: saveLessons },
+  scenarios, judge, agent,
+})
+```
+
 ### Improve a knowledge base
 
-`runKnowledgeImprovementJob` is the runtime-owned front door for KB, wiki, memory-backed, and RAG improvement jobs. It creates a candidate copy, runs supervised agents against it, checks readiness through `@tangle-network/agent-knowledge`, measures spend and timing, and promotes only when the candidate passes.
+`runKnowledgeImprovementJob` is the runtime-owned front door for KB, wiki, memory-backed, and RAG improvement jobs. It creates a candidate copy, runs supervised agents against it, checks readiness through `@tangle-network/agent-knowledge`, measures spend and timing, and promotes only when the candidate passes. Use `improve(..., { surface: 'memory' })` for the agent's curated lesson document; use this job for source, retrieval, and knowledge-store changes.
 
 ```ts
 import { runKnowledgeImprovementJob } from '@tangle-network/agent-runtime/knowledge'
diff --git a/docs/agent-optimization-map.md b/docs/agent-optimization-map.md
deleted file mode 100644
index 09c1b29c..00000000
--- a/docs/agent-optimization-map.md
+++ /dev/null
@@ -1,91 +0,0 @@
-# The Agent Optimization Map
-
-**What this document is.**
-The one-page truth about optimizing agents on this substrate: every lever an `AgentProfile` exposes, what machinery exists to optimize or create each one, what is actually wired to the paved path, and what has been proven on live infrastructure.
-Written to be read by humans and by agents configuring an improvement run.
-Audited 2026-07-06 against `agent-interface`, `agent-eval`, `agent-runtime`, `agent-knowledge`, and the supervisor-lab live campaigns (runs 1–6).
-
-## The mental model in three sentences
-
-An **AgentProfile is the complete specification of an agent**: its brain-instructions, capabilities, and starting world.
-**`improve()` is the one verb** that optimizes any single lever of that profile against an executable judge, with a statistical held-out gate deciding promotion.
-Everything else in the stack is either a **proposer** (something that generates candidate lever-values), a **judge** (something that scores results), or a **loop** (something that runs proposers against judges until the gate ships or holds).
-
-## Lever inventory — what an AgentProfile actually carries
-
-| Lever (your words) | Profile field | In the type? |
-|---|---|---|
-| system prompt | `prompt.systemPrompt` (+ per-mode `modes`) | yes |
-| skills | `resources.skills` (SKILL.md packages) | yes |
-| tools | `tools` (enable/disable map) + `resources.tools` (file-based tool defs) | yes |
-| MCP | `mcp` (server map) + `connections` | yes |
-| knowledge base | no first-class field — lives as `resources.files` content or app-side (agent-knowledge KB store) | partial |
-| resources | `resources.{instructions,agents,commands}` | yes |
-| starting files | `resources.files` (materialized into the workspace before execution) | yes |
-| full sandbox/VM state | backend/image choice + `extensions` (non-portable) + `resources.files` | partial — files yes, image/toolchain is the tcloud backend's, not the profile's |
-
-Also in the type and often forgotten: `subagents` (native sub-agents), `permissions`, `model` hints, `hooks`.
-
-## Optimization matrix — lever × machinery × wiring × proof
-
-| Surface | Baseline/apply plumbing | Proposer that exists | Reachable from `improve()`? | Proven live? |
-|---|---|---|---|---|
-| `prompt` | yes | `gepaProposer` (Pareto frontier + crossover-merge + reflective mutation) | **yes (default)** | **yes — supervisor-lab runs 3–6, ~1,700 sandbox cells** |
-| `skills` | yes | `skillOptProposer` (skill-document patching) | **yes (default)** | no live run anywhere yet |
-| `tools` | yes (JSON string surface) | none wired; `parameterSweepProposer` is shape-compatible for toggle/config sweeps | no — fails loud, needs `generator` | no |
-| `mcp` | yes (JSON string surface) | none | no — fails loud | no |
-| `hooks` | yes (JSON string surface) | none | no — fails loud | no |
-| `code` (tools/harness/anything in a repo) | yes | `improvementDriver` + `agenticGenerator` (real coding harness per candidate worktree, verify-gated) | **yes — `code: { repoRoot }` facade (#480)** | offline test only; live milestone pending |
-| `resources.files` (starting world) | **no surface** | none | no | no |
-| knowledge base | **no surface** | agent-knowledge loops exist (below) but nothing writes back into a profile | no | no |
-
-**The one-sentence verdict: every lever you believe in is representable; two are optimizable-and-proven (prompt today, code as of #480); skills is one command away; tools/mcp/hooks are string surfaces awaiting a config proposer; files/knowledge aren't surfaces at all yet.**
-
-## The proposer zoo — what each one is for (plain words)
-
-| Proposer | What it does | Wired to paved path? |
-|---|---|---|
-| `gepaProposer` | evolves prompt text: keeps a frontier of variants that each win somewhere, merges them, mutates with evidence from best/worst trials | yes (prompt default) |
-| `skillOptProposer` | same idea for skill documents (structured patches) | yes (skills default) |
-| `fapoProposer` | **the meta-policy**: attribute each failure to a level, propose ONE scoped change, escalate prompt → parameters → structure only when the cheaper level is exhausted — with pluggable per-level generators | **no** — exported, zero consumers |
-| `traceAnalystProposer` | turns trace-analyst findings (an Ax agent that reads execution traces with tools) into candidates | **no** — exported, zero consumers |
-| `aceProposer` | append-only "playbook" curator: accumulates provenance-tagged lessons without ever summarizing old ones away (anti context-collapse) | no |
-| `memoryCurationProposer` | the dedup-and-replace contrast to ACE | no |
-| `parameterSweepProposer` | sweeps numeric/config parameters | no |
-| `haloProposer` | hierarchical analyst-driven optimization | no |
-
-Since #480/#310/#311 the paved path also has: per-generation **failure distiller** (the proposer automatically sees each generation's worst cells + judge reasons), **power preflight** (minimum detectable lift computed from baseline cells; structurally-hopeless budgets warn), durable `runDir`, and the statistical holdout gate everywhere (the point-estimate gate was folded away).
-
-## Creation loops — making new things, not just mutating strings
-
-What you asked for → what exists:
-
-- **"Create new skills by mining traces"** — components all exist, composition does not: `trace-analyst` (reads traces, emits findings) → `skillOptProposer`/`aceProposer` (turn findings into skill-document content) → `improve(surface: 'skills')` (prove them). Nobody has connected the three. One composition, high value.
-- **"Auto-research loops that write code"** — exists as of #480: `improve(surface: 'code', code: { repoRoot, verify })` runs a real coding harness per candidate in a worktree, gated by any executable judge. "Create a new tool" = point it at the repo where tools live.
-- **"Knowledge discovery / find relevant OSS libraries / find MCPs"** — the research machinery exists in `agent-knowledge` (`runVerifiedResearchLoop`: two-agent verified research with claim-grounding; `discovery`; collection drivers) and is consumed by physim/legal/tax products — but **no bridge writes research results into an AgentProfile lever**. The missing primitive is small: a `knowledge → resources.files/skills` materializer, after which "research the ecosystem, propose an MCP server for the profile, prove it on the benchmark" is a normal `improve()` run over the `mcp` surface.
-- **"Full sandbox initial state as a lever"** — `resources.files` covers workspace seeding; making it an `ImproveSurface` (candidate = a file-set, judge = task performance) is unbuilt but fits the existing surface contract exactly.
-
-## The three gaps that matter (everything else is garnish)
-
-1. **Reachability**: the zoo (FAPO, trace-analyst, ACE, sweep) needs one dial — `improve(..., { proposer: 'fapo' })` — instead of hand-assembly. FAPO especially: it *is* the "optimize everything, escalate sensibly" policy this document describes, already grounded in the paper, already pluggable.
-2. **The knowledge bridge**: research loops produce verified knowledge; profiles can carry files/skills; nothing connects them.
-3. **This document didn't exist.** Now it does; keep it honest — every row's "proven live" column only flips with a linked run.
-
-## Recipes (copy-paste truth, today)
-
-```ts
-// Optimize a prompt against any executable judge (proven, runs 3–6):
-await improve(profile, findings, { surface: 'prompt', scenarios, judge, agent, runDir })
-
-// Optimize the skill documents:
-await improve(profile, findings, { surface: 'skills', scenarios, judge, agent, runDir })
-
-// Evolve CODE (tools, harnesses, algorithms) with a real coding agent per candidate:
-await improve(profile, findings, {
-  surface: 'code',
-  code: { repoRoot, verify: commandVerifier({ command: 'pnpm', args: ['test'] }) },
-  scenarios, judge, agent, runDir,
-})
-
-// Every result now carries result.power — read it BEFORE buying a bigger search.
-```
diff --git a/docs/api/index.md b/docs/api/index.md
index 249a21b2..1d0c1b39 100644
--- a/docs/api/index.md
+++ b/docs/api/index.md
@@ -5688,7 +5688,7 @@ Test seam — inject the worktree-dirty check (defaults to `git status`).
 
 ### ImproveSkillsOptions
 
-Defined in: [improvement/improve.ts:143](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L143)
+Defined in: [improvement/improve.ts:151](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L151)
 
 #### Properties
 
@@ -5696,15 +5696,15 @@ Defined in: [improvement/improve.ts:143](https://github.com/tangle-network/agent
 
 > **document**: `string`
 
-Defined in: [improvement/improve.ts:145](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L145)
+Defined in: [improvement/improve.ts:153](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L153)
 
 The skill document's current text — the baseline `skillOptProposer` patches.
 
 ##### writeBack?
 
-> `optional` **writeBack?**: (`winnerDocument`) => `void`
+> `optional` **writeBack?**: (`winnerDocument`) => `void` \| `Promise`\<`void`\>
 
-Defined in: [improvement/improve.ts:149](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L149)
+Defined in: [improvement/improve.ts:157](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L157)
 
 Persist the shipped winner document (write the file the profile ref points at).
  Called only on a ship verdict. When omitted, the winner is still returned in
@@ -5718,13 +5718,47 @@ Persist the shipped winner document (write the file the profile ref points at).
 
 ###### Returns
 
-`void`
+`void` \| `Promise`\<`void`\>
+
+***
+
+### ImproveMemoryOptions
+
+Defined in: [improvement/improve.ts:160](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L160)
+
+#### Properties
+
+##### document
+
+> **document**: `string`
+
+Defined in: [improvement/improve.ts:162](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L162)
+
+Current durable memory text used as the measured baseline.
+
+##### writeBack?
+
+> `optional` **writeBack?**: (`winnerDocument`) => `void` \| `Promise`\<`void`\>
+
+Defined in: [improvement/improve.ts:164](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L164)
+
+Persist the promoted memory document. Never called on hold or error.
+
+###### Parameters
+
+###### winnerDocument
+
+`string`
+
+###### Returns
+
+`void` \| `Promise`\<`void`\>
 
 ***
 
 ### ImproveCodeOptions
 
-Defined in: [improvement/improve.ts:152](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L152)
+Defined in: [improvement/improve.ts:167](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L167)
 
 #### Properties
 
@@ -5732,7 +5766,7 @@ Defined in: [improvement/improve.ts:152](https://github.com/tangle-network/agent
 
 > **repoRoot**: `string`
 
-Defined in: [improvement/improve.ts:154](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L154)
+Defined in: [improvement/improve.ts:169](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L169)
 
 Repo root candidate worktrees fork from.
 
@@ -5740,7 +5774,7 @@ Repo root candidate worktrees fork from.
 
 > `optional` **baseRef?**: `string`
 
-Defined in: [improvement/improve.ts:156](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L156)
+Defined in: [improvement/improve.ts:171](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L171)
 
 Base ref candidates fork from. Default `main`.
 
@@ -5748,7 +5782,7 @@ Base ref candidates fork from. Default `main`.
 
 > `optional` **worktreeDir?**: `string`
 
-Defined in: [improvement/improve.ts:158](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L158)
+Defined in: [improvement/improve.ts:173](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L173)
 
 Directory worktrees are created under. Default `<repoRoot>/.worktrees`.
 
@@ -5756,7 +5790,7 @@ Directory worktrees are created under. Default `<repoRoot>/.worktrees`.
 
 > `optional` **harness?**: [`LocalHarness`](mcp.md#localharness)
 
-Defined in: [improvement/improve.ts:160](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L160)
+Defined in: [improvement/improve.ts:175](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L175)
 
 Coding harness the agentic generator runs in each worktree. Default `claude`.
 
@@ -5764,7 +5798,7 @@ Coding harness the agentic generator runs in each worktree. Default `claude`.
 
 > `optional` **verify?**: [`Verifier`](#verifier)
 
-Defined in: [improvement/improve.ts:163](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L163)
+Defined in: [improvement/improve.ts:178](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L178)
 
 Verify a candidate worktree before it becomes a measurable surface; failures
  feed the next shot (see `agenticGenerator.verify` / `commandVerifier`).
@@ -5773,7 +5807,7 @@ Verify a candidate worktree before it becomes a measurable surface; failures
 
 > `optional` **timeoutMs?**: `number`
 
-Defined in: [improvement/improve.ts:165](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L165)
+Defined in: [improvement/improve.ts:180](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L180)
 
 Per-shot wall-clock timeout for the harness (ms).
 
@@ -5781,7 +5815,7 @@ Per-shot wall-clock timeout for the harness (ms).
 
 > `optional` **generator?**: [`CandidateGenerator`](#candidategenerator)
 
-Defined in: [improvement/improve.ts:168](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L168)
+Defined in: [improvement/improve.ts:183](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L183)
 
 Byte-producer override — the test seam and the escape hatch for custom
  candidate production. When set, `harness`/`verify`/`timeoutMs` are unused.
@@ -5790,7 +5824,7 @@ Byte-producer override — the test seam and the escape hatch for custom
 
 ### ImproveResult
 
-Defined in: [improvement/improve.ts:171](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L171)
+Defined in: [improvement/improve.ts:186](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L186)
 
 #### Type Parameters
 
@@ -5808,7 +5842,7 @@ Defined in: [improvement/improve.ts:171](https://github.com/tangle-network/agent
 
 > **profile**: `AgentProfile`
 
-Defined in: [improvement/improve.ts:174](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L174)
+Defined in: [improvement/improve.ts:189](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L189)
 
 The profile after improvement: the winner surface applied back into the
  matching field when the gate shipped, else the input profile unchanged.
@@ -5817,7 +5851,7 @@ The profile after improvement: the winner surface applied back into the
 
 > **shipped**: `boolean`
 
-Defined in: [improvement/improve.ts:176](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L176)
+Defined in: [improvement/improve.ts:191](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L191)
 
 True when `gateDecision === 'ship'`.
 
@@ -5825,7 +5859,7 @@ True when `gateDecision === 'ship'`.
 
 > **lift**: `number`
 
-Defined in: [improvement/improve.ts:178](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L178)
+Defined in: [improvement/improve.ts:193](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L193)
 
 Held-out lift (`winner − baseline` composite).
 
@@ -5833,7 +5867,7 @@ Held-out lift (`winner − baseline` composite).
 
 > **gateDecision**: `"ship"` \| `"hold"` \| `"need_more_work"` \| `"model_ceiling"` \| `"arch_ceiling"`
 
-Defined in: [improvement/improve.ts:180](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L180)
+Defined in: [improvement/improve.ts:195](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L195)
 
 The five-valued gate verdict from `selfImprove`.
 
@@ -5841,7 +5875,7 @@ The five-valued gate verdict from `selfImprove`.
 
 > **raw**: `SelfImproveResult`\<`TScenario`, `TArtifact`\>
 
-Defined in: [improvement/improve.ts:182](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L182)
+Defined in: [improvement/improve.ts:197](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L197)
 
 Full `selfImprove` result for advanced inspection.
 
@@ -10632,9 +10666,9 @@ Verifies the edited worktree. Sync or async; throws only on a setup fault
 
 ### ImproveSurface
 
-> **ImproveSurface** = `"prompt"` \| `"skills"` \| `"tools"` \| `"mcp"` \| `"hooks"` \| `"subagents"` \| `"workflow"` \| `"agent-profile"` \| `"code"` \| `"rollout-policy"`
+> **ImproveSurface** = `"prompt"` \| `"skills"` \| `"tools"` \| `"mcp"` \| `"hooks"` \| `"subagents"` \| `"workflow"` \| `"agent-profile"` \| `"memory"` \| `"code"` \| `"rollout-policy"`
 
-Defined in: [improvement/improve.ts:75](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L75)
+Defined in: [improvement/improve.ts:78](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L78)
 
 The agent-profile lever `improve` optimizes. Mirrors the AgentProfile-law
  profile levers; `code` is the implementation-tier surface, `rollout-policy`
@@ -10647,7 +10681,7 @@ The agent-profile lever `improve` optimizes. Mirrors the AgentProfile-law
 
 > **ImproveOptions**\<`TScenario`, `TArtifact`\> = `Omit`\<`SelfImproveOptions`\<`TScenario`, `TArtifact`\>, `"analyzeGeneration"` \| `"baselineSurface"` \| `"findings"` \| `"gate"` \| `"proposer"`\> & `object`
 
-Defined in: [improvement/improve.ts:87](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L87)
+Defined in: [improvement/improve.ts:91](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L91)
 
 #### Type Declaration
 
@@ -10663,8 +10697,8 @@ Which profile lever to optimize. Default `'prompt'`. Selects the default
 > `optional` **generator?**: `SurfaceProposer`
 
 The `SurfaceProposer` that mutates the surface. When unset, the facade
- picks the default for `surface` (`gepaProposer` for prompt, `skillOptProposer`
- for skills); surfaces with no default REQUIRE this (fail-loud otherwise).
+ picks the default for prompt, skills, memory, and rollout policy; surfaces
+ with no default REQUIRE this (fail-loud otherwise).
 
 ##### gate?
 
@@ -10730,6 +10764,14 @@ SKILLS-surface wiring for real skill-DOCUMENT optimization. Without this,
  shipped winner (the profile ref points at a file the caller owns). This is
  what makes skillOpt reachable through improve().
 
+##### memory?
+
+> `optional` **memory?**: [`ImproveMemoryOptions`](#improvememoryoptions)
+
+MEMORY-surface wiring for a curated durable memory document. The default
+ deterministic proposer deduplicates and ranks lessons from findings, then
+ replaces its managed block instead of growing memory without bound.
+
 ##### promotionGate?
 
 > `optional` **promotionGate?**: `SelfImproveOptions`\<`TScenario`, `TArtifact`\>\[`"gate"`\]
@@ -12592,7 +12634,7 @@ Build the starting instruction for a coder agent tasked with implementing a new
 
 > **improve**\<`TScenario`, `TArtifact`\>(`profile`, `findings`, `opts`): `Promise`\<[`ImproveResult`](#improveresult)\<`TScenario`, `TArtifact`\>\>
 
-Defined in: [improvement/improve.ts:465](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L465)
+Defined in: [improvement/improve.ts:510](https://github.com/tangle-network/agent-runtime/blob/main/src/improvement/improve.ts#L510)
 
 Run the held-out-gated self-improvement loop on ONE profile surface.
 
diff --git a/docs/api/intelligence.md b/docs/api/intelligence.md
index 84684047..4bc0930e 100644
--- a/docs/api/intelligence.md
+++ b/docs/api/intelligence.md
@@ -3564,7 +3564,7 @@ Analyze one run and produce one measured, review-only improvement proposal.
 
 > **reviewAgentImprovementProposal**(`inputProposal`, `input`): [`AgentImprovementReview`](#agentimprovementreview)
 
-Defined in: [intelligence/improvement-cycle.ts:201](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L201)
+Defined in: [intelligence/improvement-cycle.ts:208](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L208)
 
 Persist an approve/reject/change-request decision bound to one exact proposal.
 
@@ -3588,7 +3588,7 @@ Persist an approve/reject/change-request decision bound to one exact proposal.
 
 > **executeApprovedAgentCandidate**(`options`): `Promise`\<[`ExecuteApprovedAgentCandidateResult`](#executeapprovedagentcandidateresult)\>
 
-Defined in: [intelligence/improvement-cycle.ts:231](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L231)
+Defined in: [intelligence/improvement-cycle.ts:238](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L238)
 
 Verify, materialize, run, grade, and receipt only the exact approved bundle.
 
@@ -3608,7 +3608,7 @@ Verify, materialize, run, grade, and receipt only the exact approved bundle.
 
 > **verifyAgentImprovementProposal**(`input`): [`AgentImprovementProposal`](#agentimprovementproposal)
 
-Defined in: [intelligence/improvement-cycle.ts:273](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L273)
+Defined in: [intelligence/improvement-cycle.ts:280](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L280)
 
 Validate a proposal's schema, profile, sealed bundle, and canonical digest.
 
@@ -3628,7 +3628,7 @@ Validate a proposal's schema, profile, sealed bundle, and canonical digest.
 
 > **verifyAgentImprovementReview**(`input`): [`AgentImprovementReview`](#agentimprovementreview)
 
-Defined in: [intelligence/improvement-cycle.ts:317](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L317)
+Defined in: [intelligence/improvement-cycle.ts:324](https://github.com/tangle-network/agent-runtime/blob/main/src/intelligence/improvement-cycle.ts#L324)
 
 Validate a review's decision fields and canonical digest.
 
diff --git a/docs/api/primitive-catalog.md b/docs/api/primitive-catalog.md
index bc6af84a..595c4d1c 100644
--- a/docs/api/primitive-catalog.md
+++ b/docs/api/primitive-catalog.md
@@ -7,7 +7,7 @@
 
 # Primitive catalog — the never-stale anti-reinvention inventory
 
-> **GENERATED** from `@tangle-network/agent-runtime@0.92.1` and `@tangle-network/agent-eval@0.113.0` by `scripts/gen-primitive-catalog.mjs`. Do NOT hand-edit — run `pnpm run docs:api`. This is the mechanical companion to the JUDGMENT in `canonical-api.md` (§2 decision table + §1.5 AgentProfile law): that doc says WHICH primitive to reach for and what NOT to build; this catalog proves WHAT exists. Per-symbol signatures + `file:line` live in the per-module pages under `docs/api/`.
+> **GENERATED** from `@tangle-network/agent-runtime@0.93.0` and `@tangle-network/agent-eval@0.114.0` by `scripts/gen-primitive-catalog.mjs`. Do NOT hand-edit — run `pnpm run docs:api`. This is the mechanical companion to the JUDGMENT in `canonical-api.md` (§2 decision table + §1.5 AgentProfile law): that doc says WHICH primitive to reach for and what NOT to build; this catalog proves WHAT exists. Per-symbol signatures + `file:line` live in the per-module pages under `docs/api/`.
 
 ## 1. agent-runtime — own public surface
 
@@ -15,7 +15,7 @@ Every subpath this package declares in `package.json` `exports`. Reach for these
 
 ### Root — task lifecycle, conversation, RSI verbs, observability
 
-Import from `@tangle-network/agent-runtime` — 330 exports.
+Import from `@tangle-network/agent-runtime` — 331 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -212,7 +212,7 @@ Import from `@tangle-network/agent-runtime` — 330 exports.
 | `ToolLoopStopReason` | type | Why the loop stopped. `completed` = model finished naturally; `stuck-loop` = |
 | `Verifier` | type | Verifies the edited worktree. Sync or async; throws only on a setup fault |
 
-**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentAdapter`, `AgentBackendContext`, `AgentBackendInput`, `AgentCandidateBenchmarkGraderIdentity`, `AgentCandidateContainerPort`, `AgentCandidateExecutionAttemptRef`, `AgentCandidateExecutionPorts`, `AgentCandidateExecutorProfileFile`, `AgentCandidateExecutorWorkspaceFile`, `AgentCandidateExecutorWorkspaceInput`, `AgentCandidateMemoryPort`, `AgentCandidateMemoryResetResult`, `AgentCandidateModelPort`, `AgentCandidateProtectedModelActivation`, `AgentCandidateProtectedModelReservation`, `AgentCandidateProtectedModelSettlement`, `AgentCandidateProtectedRunCapture`, `AgentCandidateTaskExecution`, `AgentCandidateVerificationPorts`, `AgentExecutionBackend`, `AgenticGeneratorOptions`, `AgentKnowledgeProvider`, `AgentKnowledgeReadinessCheckOptions`, `AgentTaskContext`, `AgentTaskRunResult`, `AgentTaskSpec`, `BackendCallPolicy`, `CanonicalCandidateDocument`, `ChatTurnHooks`, `ChatTurnResult`, `ControlBudget`, `ControlEvalResult`, `ControlRunResult`, `ControlStep`, `Conversation`, `ConversationDriveState`, `ConversationJournal`, `ConversationParticipant`, `ConversationPolicy`, `ConversationResult`, `ConversationTurn`, `CreateProtectedAgentCandidateModelPortOptions`, `D1StmtLike`, `DataAcquisitionPlan`, `DelegatedLoopResult`, `DisposePreparedAgentCandidateOptions`, `EvalRunEvent`, `EvalRunGeneration`, `EvalRunsExportConfig`, `EvalRunsExportResult`, `ExecutePreparedAgentCandidateOptions`, `FileAgentCandidateExecutionClaimStoreOptions`, `HaltContext`, `HaltSignal`, `ImproveCodeOptions`, `ImprovementDriverOptions`, `ImproveResult`, `ImproveSkillsOptions`, `KnowledgeImprovementJobMeasurement`, `KnowledgeImprovementJobResult`, `KnowledgeReadinessCheckInput`, `KnowledgeReadinessReport`, `KnowledgeRequirement`, `LoopRunnerCliArgs`, `LoopRunnerCliResult`, `ManagedImprovementDriver`, `OtelAttribute`, `OtelExporter`, `OtelSpan`, `PersonaConversationResult`, `PrepareAgentCandidateExecutionOptions`, `PreparedAgentCandidateExecution`, `PreparedAgentCandidateInstruction`, `PreparedAgentCandidateLaunch`, `PreparedAgentCandidateTrace`, `RecoverExpiredAgentCandidateOptions`, `ResearchLoopResult`, `ResearchLoopRunnerOptions`, `ResolveAgentBackendOptions`, `ResolvedAgentCandidateContainer`, `ResolvedChatModel`, `RunChatTurnInput`, `RunConversationOptions`, `RunDelegatedLoopOptions`, `RunKnowledgeImprovementJobOptions`, `RunPersonaConfig`, `RunPersonaConversationOptions`, `RuntimeDecisionEvidenceRef`, `RuntimeDecisionPoint`, `RuntimeEventCollector`, `RuntimeEventOtelOptions`, `RuntimeHookContext`, `RuntimeHookErrorContext`, `RuntimeHookEvent`, `RuntimeRunHandle`, `RuntimeRunPersistenceAdapter`, `RuntimeRunRow`, `RuntimeSessionStore`, `RuntimeStreamEventCollector`, `RuntimeTelemetryOptions`, `RunToolLoopOptions`, `SanitizedKnowledgeReadinessReport`, `StreamToolLoopOptions`, `SupervisedKnowledgeUpdateInput`, `SupervisedKnowledgeUpdateOptions`, `SupervisedKnowledgeUpdateResult`, `ToolLoopResult`, `VerifiedAgentCandidate`, `VetoedFact`, `WorktreeLoopRunnerOptions`, `AgentCandidateModelGrantActivateInput`, `AgentCandidateModelGrantReserveInput`, `AgentCandidateModelGrantSettleInput`, `AgentCandidateOutputPurpose`, `AgentCandidateRetryRejection`, `AgentCandidateRunFinalization`, `AgentRuntimeEvent`, `AgentRuntimeEventSink`, `AgentTaskStatus`, `AuthSource`, `ControlDecision`, `ConversationStreamEvent`, `DelegatedLoopMode`, `DelegatedLoopRegistry`, `DelegatedLoopRunner`, `ForwardHeaderName`, `HaltPredicate`, `HaltReason`, `ImproveOptions`, `KnowledgeReadinessCheck`, `KnowledgeReadinessCheckResult`, `RuntimeDecisionKind`, `RuntimeHookTarget`, `RuntimeStreamEvent`, `StreamToolLoopYield`, `SupervisedKnowledgeUpdater`, `ToolLoopEvent`, `TurnOrder`.
+**Undocumented supporting types** (add a TSDoc line at the declaration to earn a table row): `AgentAdapter`, `AgentBackendContext`, `AgentBackendInput`, `AgentCandidateBenchmarkGraderIdentity`, `AgentCandidateContainerPort`, `AgentCandidateExecutionAttemptRef`, `AgentCandidateExecutionPorts`, `AgentCandidateExecutorProfileFile`, `AgentCandidateExecutorWorkspaceFile`, `AgentCandidateExecutorWorkspaceInput`, `AgentCandidateMemoryPort`, `AgentCandidateMemoryResetResult`, `AgentCandidateModelPort`, `AgentCandidateProtectedModelActivation`, `AgentCandidateProtectedModelReservation`, `AgentCandidateProtectedModelSettlement`, `AgentCandidateProtectedRunCapture`, `AgentCandidateTaskExecution`, `AgentCandidateVerificationPorts`, `AgentExecutionBackend`, `AgenticGeneratorOptions`, `AgentKnowledgeProvider`, `AgentKnowledgeReadinessCheckOptions`, `AgentTaskContext`, `AgentTaskRunResult`, `AgentTaskSpec`, `BackendCallPolicy`, `CanonicalCandidateDocument`, `ChatTurnHooks`, `ChatTurnResult`, `ControlBudget`, `ControlEvalResult`, `ControlRunResult`, `ControlStep`, `Conversation`, `ConversationDriveState`, `ConversationJournal`, `ConversationParticipant`, `ConversationPolicy`, `ConversationResult`, `ConversationTurn`, `CreateProtectedAgentCandidateModelPortOptions`, `D1StmtLike`, `DataAcquisitionPlan`, `DelegatedLoopResult`, `DisposePreparedAgentCandidateOptions`, `EvalRunEvent`, `EvalRunGeneration`, `EvalRunsExportConfig`, `EvalRunsExportResult`, `ExecutePreparedAgentCandidateOptions`, `FileAgentCandidateExecutionClaimStoreOptions`, `HaltContext`, `HaltSignal`, `ImproveCodeOptions`, `ImproveMemoryOptions`, `ImprovementDriverOptions`, `ImproveResult`, `ImproveSkillsOptions`, `KnowledgeImprovementJobMeasurement`, `KnowledgeImprovementJobResult`, `KnowledgeReadinessCheckInput`, `KnowledgeReadinessReport`, `KnowledgeRequirement`, `LoopRunnerCliArgs`, `LoopRunnerCliResult`, `ManagedImprovementDriver`, `OtelAttribute`, `OtelExporter`, `OtelSpan`, `PersonaConversationResult`, `PrepareAgentCandidateExecutionOptions`, `PreparedAgentCandidateExecution`, `PreparedAgentCandidateInstruction`, `PreparedAgentCandidateLaunch`, `PreparedAgentCandidateTrace`, `RecoverExpiredAgentCandidateOptions`, `ResearchLoopResult`, `ResearchLoopRunnerOptions`, `ResolveAgentBackendOptions`, `ResolvedAgentCandidateContainer`, `ResolvedChatModel`, `RunChatTurnInput`, `RunConversationOptions`, `RunDelegatedLoopOptions`, `RunKnowledgeImprovementJobOptions`, `RunPersonaConfig`, `RunPersonaConversationOptions`, `RuntimeDecisionEvidenceRef`, `RuntimeDecisionPoint`, `RuntimeEventCollector`, `RuntimeEventOtelOptions`, `RuntimeHookContext`, `RuntimeHookErrorContext`, `RuntimeHookEvent`, `RuntimeRunHandle`, `RuntimeRunPersistenceAdapter`, `RuntimeRunRow`, `RuntimeSessionStore`, `RuntimeStreamEventCollector`, `RuntimeTelemetryOptions`, `RunToolLoopOptions`, `SanitizedKnowledgeReadinessReport`, `StreamToolLoopOptions`, `SupervisedKnowledgeUpdateInput`, `SupervisedKnowledgeUpdateOptions`, `SupervisedKnowledgeUpdateResult`, `ToolLoopResult`, `VerifiedAgentCandidate`, `VetoedFact`, `WorktreeLoopRunnerOptions`, `AgentCandidateModelGrantActivateInput`, `AgentCandidateModelGrantReserveInput`, `AgentCandidateModelGrantSettleInput`, `AgentCandidateOutputPurpose`, `AgentCandidateRetryRejection`, `AgentCandidateRunFinalization`, `AgentRuntimeEvent`, `AgentRuntimeEventSink`, `AgentTaskStatus`, `AuthSource`, `ControlDecision`, `ConversationStreamEvent`, `DelegatedLoopMode`, `DelegatedLoopRegistry`, `DelegatedLoopRunner`, `ForwardHeaderName`, `HaltPredicate`, `HaltReason`, `ImproveOptions`, `KnowledgeReadinessCheck`, `KnowledgeReadinessCheckResult`, `RuntimeDecisionKind`, `RuntimeHookTarget`, `RuntimeStreamEvent`, `StreamToolLoopYield`, `SupervisedKnowledgeUpdater`, `ToolLoopEvent`, `TurnOrder`.
 
 ### Vertical agent — manifest + improvement adapter
 
@@ -1105,7 +1105,7 @@ Import from `@tangle-network/agent-eval` — 10 exports.
 
 ### STATISTICS — significance, intervals, effect size
 
-Import from `@tangle-network/agent-eval` — 50 exports.
+Import from `@tangle-network/agent-eval` — 51 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -1128,6 +1128,7 @@ Import from `@tangle-network/agent-eval` — 50 exports.
 | `pairedEvalueSequence` | function | Run the paired e-value sequence over an in-order delta stream. |
 | `pairedMde` | function | Minimum detectable paired effect (standardised units) for a target paired |
 | `pairedRiskDifference` | function | Paired risk difference (the effect-size companion to {@link mcnemar}): the |
+| `pairedSignTest` | function | Exact one-sided sign test over paired differences. |
 | `pairedTTest` | function | Paired t-test — before/after measurements on the SAME items. |
 | `partialCredit` | function | Partial credit: returns 0-1 ratio of current toward target |
 | `passAtK` | function | Unbiased pass@k for code generation (Chen et al. 2021, "Evaluating Large |
diff --git a/docs/canonical-api.md b/docs/canonical-api.md
index 89d0c72d..c2056e0a 100644
--- a/docs/canonical-api.md
+++ b/docs/canonical-api.md
@@ -2,7 +2,7 @@
 
 <!-- This doc is the JUDGMENT layer: the mental model (§1), the AgentProfile law (§1.5), and the anti-reinvention decision table (§2) — WHICH primitive to reach for and what NOT to build. The export INVENTORY (WHAT exists) and per-symbol signatures + `file:line` are GENERATED into `docs/api/` (TypeDoc + `scripts/gen-primitive-catalog.mjs`, do NOT hand-edit) — that is the mechanical reference: `docs/api/primitive-catalog.md` is the never-stale list of every primitive to reuse. The freshness gate (`pnpm docs:freshness`) FAILS CI if a version pin, a cited `file:line`, a decision-table symbol, or the generated catalog drifts from source — see `docs/MAINTAINING.md`. Keep this file the small, hand-curated spine; never re-list the inventory here — point at the catalog. -->
 
-> **Version 0.92.1.** The export inventory + per-symbol signatures live in the generated `docs/api/` reference: **`docs/api/primitive-catalog.md`** is the never-stale, grouped list of every primitive to reuse (own surface + the agent-eval judge / authenticity / verification / statistics / campaign / token-usage surfaces), with each one's import path and one-line summary read live from source; the per-module pages hold the full signatures. The pinned substrate is agent-eval `>=0.113.0 <1.0.0`; the sandbox substrate that materializes profiles into harness shapes is `@tangle-network/sandbox` (peer `>=0.8.0 <1.0.0`). The neutral contract types (`AgentProfile`, `AgentProfileMcpServer`, `HarnessType`, `ReasoningEffort`, `Part`/`ToolPart`/`ToolState`, plus environment-provider types) are owned by **`@tangle-network/agent-interface`** (peer `>=0.25.0 <1.0.0`) — the single source of truth. Substrate primitives are re-exported through `@tangle-network/agent-eval/contract` (or `/campaign`), not local to this package — the catalog's §2 shows exactly which subpath each lives under.
+> **Version 0.93.0.** The export inventory + per-symbol signatures live in the generated `docs/api/` reference: **`docs/api/primitive-catalog.md`** is the never-stale, grouped list of every primitive to reuse (own surface + the agent-eval judge / authenticity / verification / statistics / campaign / token-usage surfaces), with each one's import path and one-line summary read live from source; the per-module pages hold the full signatures. The pinned substrate is agent-eval `>=0.114.0 <1.0.0`; the sandbox substrate that materializes profiles into harness shapes is `@tangle-network/sandbox` (peer `>=0.8.0 <1.0.0`). The neutral contract types (`AgentProfile`, `AgentProfileMcpServer`, `HarnessType`, `ReasoningEffort`, `Part`/`ToolPart`/`ToolState`, plus environment-provider types) are owned by **`@tangle-network/agent-interface`** (peer `>=0.25.0 <1.0.0`) — the single source of truth. Substrate primitives are re-exported through `@tangle-network/agent-eval/contract` (or `/campaign`), not local to this package — the catalog's §2 shows exactly which subpath each lives under.
 >
 > **`./loops` is the runtime barrel** — `package.json` maps it to `src/runtime/index.ts`. Everything below labelled `/loops` is the recursive-atom + loop-kernel surface.
 >
@@ -86,7 +86,7 @@ A general "loop" primitive is the single most common modelling error in this rep
 | Pick the **chat backend an in-process turn runs on** (`router`/`tcloud`/`cli-bridge`/`sandbox`) from a product flag | `resolveAgentBackend({ backend })` — root `.` | the copy-pasted `backend-name → createOpenAICompatibleBackend` branch every eval product hand-rolled (the copies drift) |
 | Pick / register a leaf backend, or bring your own agent | `createExecutor({ backend })` / `createExecutorRegistry()` / implement `Executor` — `/loops` | a per-vendor adapter or closed `inline\|sandbox\|cli` switch (won't report through the `UsageEvent` channel) |
 | Evolve a **prompt/string** surface | `gepaProposer({ llm, model, target })` (default inside `selfImprove`; the skill-surface twin is `skillOptProposer`, same source) — `agent-eval/campaign` | a hand-rolled prompt-mutation reflection loop with its own Pareto bookkeeping |
-| Self-improve a profile (one pluggable verb) — START HERE (self-improvement) | `improve(profile, findings, { surface, gate })` — root `.`; prompt and skill-document surfaces have built-in generators, tools/MCP/hooks/subagents/workflow/whole-profile take an explicit generator, rollout policy is deterministic, and code gets isolated incumbent/candidate worktrees | a bespoke optimize loop, calling `selfImprove`/a skill optimizer directly for the common case, or comparing code against an empty/string stand-in |
+| Self-improve a profile (one pluggable verb) — START HERE (self-improvement) | `improve(profile, findings, { surface, gate })` — root `.`; prompt, skill-document, and curated-memory surfaces have built-in generators, tools/MCP/hooks/subagents/workflow/whole-profile take an explicit generator, rollout policy is deterministic, and code gets isolated incumbent/candidate worktrees | a bespoke optimize loop, calling `selfImprove`/a skill or memory optimizer directly for the common case, or comparing code against an empty/string stand-in |
 | Measure **one profile artifact's marginal lift** (with-vs-without, score+cost) / catalog artifacts | `measureMarginalLift(...)` / `ArtifactRegistry` (`applyArtifact` is the one `ArtifactKind`→`AgentProfile`-field bridge) — `/lifecycle` | a hand-rolled with/without ablation loop, or a per-kind `if kind==='skill'…` profile-field switch |
 | Run the **whole artifact lifecycle** — generate→measure→promote→store→compose, then drift-watch/dedupe the live set — over ANY profile surface (skill/prompt/tool/MCP) | `runLifecycle({ baseline, generators, evalRunner, gate })` then `composeProfile(registry, base, query)`; maintain with `driftWatch(...)` / `dedupeArtifacts(...)` — `/lifecycle` | a per-surface improve loop, a hand-rolled promote→compose step, or re-running `measureMarginalLift` without the registry/gate spine. The ONLY per-surface code is a thin `CandidateGenerator` (`skillGenerator` distills, `promptGenerator`/`buildableGenerator` for the rest) |
 | Run the self-improvement loop with full substrate control | `selfImprove({ agent, scenarios, judge, baselineSurface })` — `agent-eval/contract` | a bespoke optimize loop or a parallel skill-optimizer |
diff --git a/package.json b/package.json
index 2b9a0d5d..ba60e5aa 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-runtime",
-  "version": "0.92.1",
+  "version": "0.93.0",
   "description": "Shared task-lifecycle skeleton for agents: a recursive loop kernel for chat turns, one-shot tasks, and multi-attempt loops, with trace capture and eval-gated self-improvement. Domain behavior lives in adapters; scoring and ship-gates in @tangle-network/agent-eval.",
   "homepage": "https://github.com/tangle-network/agent-runtime#readme",
   "repository": {
@@ -104,7 +104,7 @@
   },
   "devDependencies": {
     "@biomejs/biome": "^2.4.15",
-    "@tangle-network/agent-eval": "^0.113.0",
+    "@tangle-network/agent-eval": "^0.114.0",
     "@tangle-network/agent-interface": "^0.25.0",
     "@tangle-network/sandbox": "^0.9.7",
     "@types/node": "^25.9.3",
@@ -134,7 +134,7 @@
   "license": "MIT",
   "packageManager": "pnpm@10.28.0",
   "peerDependencies": {
-    "@tangle-network/agent-eval": ">=0.113.0 <1.0.0",
+    "@tangle-network/agent-eval": ">=0.114.0 <1.0.0",
     "@tangle-network/agent-interface": ">=0.25.0 <1.0.0",
     "@tangle-network/sandbox": ">=0.8.0 <1.0.0",
     "playwright": "^1.40.0"
@@ -149,6 +149,6 @@
   },
   "dependencies": {
     "@tangle-network/agent-knowledge": "^1.11.2",
-    "@tangle-network/agent-profile-materialize": "0.3.1"
+    "@tangle-network/agent-profile-materialize": "0.3.2"
   }
 }
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index 38d4b547..e771d8d9 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -12,15 +12,15 @@ importers:
         specifier: ^1.11.2
         version: 1.11.2(typescript@5.9.3)
       '@tangle-network/agent-profile-materialize':
-        specifier: 0.3.1
-        version: 0.3.1
+        specifier: 0.3.2
+        version: 0.3.2
     devDependencies:
       '@biomejs/biome':
         specifier: ^2.4.15
         version: 2.4.15
       '@tangle-network/agent-eval':
-        specifier: ^0.113.0
-        version: 0.113.0(typescript@5.9.3)
+        specifier: ^0.114.0
+        version: 0.114.0(typescript@5.9.3)
       '@tangle-network/agent-interface':
         specifier: ^0.25.0
         version: 0.25.0
@@ -648,6 +648,11 @@ packages:
     engines: {node: '>=20'}
     hasBin: true
 
+  '@tangle-network/agent-eval@0.114.0':
+    resolution: {integrity: sha512-pIZCLkKsHeKDpLOpRXRhnc2HaREcHduHm4u3jw0nN73B1BWe6Jaqh2wlNHTPnPUA1qQl5sCl1K6XF1p9R7eTZw==}
+    engines: {node: '>=20'}
+    hasBin: true
+
   '@tangle-network/agent-interface@0.13.0':
     resolution: {integrity: sha512-CeTPGRLoXqpt0h+BCyFgZPkfU1zyRpWmqfD+85i/uk+uvbqxkfI+JprfKVf3tBsQuCgJPSjPt5qjdW8n3h2BVg==}
 
@@ -665,8 +670,8 @@ packages:
     engines: {node: '>=20'}
     hasBin: true
 
-  '@tangle-network/agent-profile-materialize@0.3.1':
-    resolution: {integrity: sha512-yA/DaxmC+DsHdPl00iovR3tg80lQY1ukkHgGViLtzFYY7O46yZR9Ykhw/tMGTFYWFi5bzqjmr0NShSxdNmSzcQ==}
+  '@tangle-network/agent-profile-materialize@0.3.2':
+    resolution: {integrity: sha512-jCj1Hc/brPQ5eS7or9A+Klv0FAZcGtyFr7kx2cKfk+wCj0/4Di3k1pMjayrmRNgI/7Oc47gt6/t+qvdtZxBkAw==}
 
   '@tangle-network/sandbox@0.9.7':
     resolution: {integrity: sha512-9pCwJ5MlF7RUpp0AQKQDFyR0yu+E0udEhWkqhrlb/RuoJxlt72zVPuzO4FnMb1MZTkfjStmomC3k5xQyqi1YSA==}
@@ -1640,6 +1645,24 @@ snapshots:
       - typescript
       - utf-8-validate
 
+  '@tangle-network/agent-eval@0.114.0(typescript@5.9.3)':
+    dependencies:
+      '@asteasolutions/zod-to-openapi': 8.5.0(zod@4.4.3)
+      '@ax-llm/ax': 19.0.45(zod@4.4.3)
+      '@hono/node-server': 2.0.8(hono@4.12.28)
+      '@tangle-network/agent-interface': 0.22.0
+      '@tangle-network/tcloud': 0.4.14(typescript@5.9.3)(zod@4.4.3)
+      hono: 4.12.28
+      zod: 4.4.3
+    transitivePeerDependencies:
+      - '@mastra/core'
+      - '@modelcontextprotocol/sdk'
+      - ai
+      - bufferutil
+      - openai
+      - typescript
+      - utf-8-validate
+
   '@tangle-network/agent-interface@0.13.0':
     dependencies:
       zod: 4.4.3
@@ -1669,9 +1692,9 @@ snapshots:
       - typescript
       - utf-8-validate
 
-  '@tangle-network/agent-profile-materialize@0.3.1':
+  '@tangle-network/agent-profile-materialize@0.3.2':
     dependencies:
-      '@tangle-network/agent-interface': 0.22.0
+      '@tangle-network/agent-interface': 0.25.0
 
   '@tangle-network/sandbox@0.9.7(viem@2.54.6(typescript@5.9.3)(zod@4.4.3))':
     dependencies:
diff --git a/skills/build-with-agent-runtime/SKILL.md b/skills/build-with-agent-runtime/SKILL.md
index 8b1d51c7..0761921c 100644
--- a/skills/build-with-agent-runtime/SKILL.md
+++ b/skills/build-with-agent-runtime/SKILL.md
@@ -93,7 +93,7 @@ to its native default (`HARNESS_NATIVE_MODEL`) — never silently dropped.
 | **Sandbox coding rollout** (fresh box/round, or persistent+resume) | `runLoop(options)` / `openSandboxRun(client, opts, deliverable)` — `/loops` | canonical-api §3.1 |
 | **Optimize a CODE surface** in a gated loop | `improvementDriver({ worktree, generator })` — root `.` | canonical-api §3.4 |
 | **Freeze a measured profile/diff + code surface for execution** | `buildAgentCandidateBundle(...)` then `verifyAgentCandidateBundle(...)` — `/candidate-execution` | canonical-api §2 |
-| **Optimize any agent/code surface** (one call) — START HERE | `improve(profile, findings, { surface, gate })` — root `.`; prompt and skill-document surfaces have built-in proposers, config/whole-profile surfaces accept a proposer, rollout policy enumerates bounded variants, and code gets isolated incumbent/candidate worktrees; drop to `selfImprove({ agent, scenarios, judge, baselineSurface })` only for lower-level control | canonical-api §3.4 |
+| **Optimize any agent/code surface** (one call) — START HERE | `improve(profile, findings, { surface, gate })` — root `.`; prompt, skill-document, and curated-memory surfaces have built-in proposers, config/whole-profile surfaces accept a proposer, rollout policy enumerates bounded variants, and code gets isolated incumbent/candidate worktrees; drop to `selfImprove({ agent, scenarios, judge, baselineSurface })` only for lower-level control | canonical-api §3.4 |
 | **Gate: ship/hold a candidate** (campaign ctx) | `defaultProductionGate` / `heldOutGate` / `composeGate` — `agent-eval/contract`; `neutralizationGate` (footprint-matched PLACEBO gate — proves a held-out lift is CONTENT, not added prompt/mount footprint) — `agent-eval/campaign` | canonical-api §3.4 |
 | **Gate: ship/hold from a `BenchmarkReport`** (per-task cells) | `promotionGate({ report, incumbent, candidate })` — `/loops` | canonical-api §3.4 |
 | **Run the full multi-generation flywheel + certify** | `runStrategyEvolution(config)` — `/loops` | canonical-api §3.4 |
diff --git a/src/improvement/improve.test.ts b/src/improvement/improve.test.ts
index 8c030407..c3f0e456 100644
--- a/src/improvement/improve.test.ts
+++ b/src/improvement/improve.test.ts
@@ -2,8 +2,8 @@
  * `improve()` default-proposer resolution proof.
  *
  * The regression this guards: `improve()` maps each surface to a default
- * `SurfaceProposer` — `prompt → gepaProposer`, `skills → skillOptProposer`.
- * Both proposers are factories exported from `@tangle-network/agent-eval/campaign`.
+ * `SurfaceProposer` — `prompt → gepaProposer`, `skills → skillOptProposer`,
+ * `memory → memoryCurationProposer`.
  * If either import resolves to `undefined` (a substrate export drift), the facade
  * does not fail at module load — it fails at CALL time, the first time a caller
  * names that surface. So a green typecheck is not enough; this test drives the
@@ -17,7 +17,13 @@
  * sees a real backend rather than a silent-zero stub.
  */
 
-import type { DispatchContext, JudgeConfig, Scenario } from '@tangle-network/agent-eval/contract'
+import type {
+  CodeSurface,
+  DispatchContext,
+  JudgeConfig,
+  Scenario,
+  SurfaceProposer,
+} from '@tangle-network/agent-eval/contract'
 import type { AgentProfile } from '@tangle-network/agent-interface'
 import { describe, expect, it } from 'vitest'
 import { ConfigError } from '../errors'
@@ -117,6 +123,145 @@ describe('improve() — default proposer resolution (substrate export drift guar
     ).rejects.toThrow(/requires opts\.skills\.document/)
   })
 
+  it("surface 'memory' resolves memoryCurationProposer and requires a real baseline", async () => {
+    const result = await improve(promptProfile(), [], {
+      surface: 'memory',
+      gate: 'none',
+      scenarios,
+      judge,
+      agent: stubAgent,
+      memory: { document: '# Durable memory\n' },
+    })
+
+    expect(result.gateDecision).toBe('hold')
+    expect(result.shipped).toBe(false)
+    await expect(
+      improve(promptProfile(), [], {
+        surface: 'memory',
+        gate: 'none',
+        scenarios,
+        judge,
+        agent: stubAgent,
+      }),
+    ).rejects.toThrow(/requires opts\.memory\.document/)
+  })
+
+  it("surface 'memory' curates seed findings and awaits persistence of the promoted document", async () => {
+    const baseline = '# Durable memory\n'
+    let writtenBack: string | null = null
+    const result = await improve(promptProfile(), [{ claim: 'improved verification lesson' }], {
+      surface: 'memory',
+      scenarios,
+      judge: improvementJudge,
+      agent: stubAgent,
+      memory: {
+        document: baseline,
+        writeBack: async (winner) => {
+          await Promise.resolve()
+          writtenBack = winner
+        },
+      },
+      promotionGate: {
+        name: 'test-ship',
+        decide: async () => ({
+          decision: 'ship',
+          reasons: ['test candidate'],
+          contributingGates: [],
+        }),
+      },
+      budget: { generations: 1, populationSize: 1, holdoutFraction: 0.25 },
+    })
+
+    expect(result.shipped).toBe(true)
+    expect(writtenBack).toContain('improved verification lesson')
+    expect(result.profile).toEqual(promptProfile())
+  })
+
+  it("surface 'memory' rejects a shipped non-text winner before persistence", async () => {
+    let writeCalls = 0
+    const nonTextSurface: CodeSurface = {
+      kind: 'code',
+      worktreeRef: 'not-a-memory-document',
+      baseRef: 'main',
+      baseCommit: 'a'.repeat(40),
+      baseTree: 'b'.repeat(40),
+      candidateCommit: 'c'.repeat(40),
+      candidateTree: 'd'.repeat(40),
+      patch: {
+        format: 'git-diff-binary',
+        sha256: `sha256:${'e'.repeat(64)}`,
+        byteLength: 1,
+      },
+    }
+    const surfaceKindJudge: JudgeConfig<{ text: string }, Scenario> = {
+      name: 'surface-kind-judge',
+      dimensions: [{ key: 'q', description: 'candidate is code-tier' }],
+      score: ({ artifact }) => {
+        const score = artifact.text === 'code' ? 1 : 0
+        return { dimensions: { q: score }, composite: score, notes: '' }
+      },
+    }
+    const malformedMemory: SurfaceProposer = {
+      kind: 'malformed-memory',
+      propose: async () => [
+        {
+          surface: nonTextSurface,
+          label: 'wrong-tier',
+          rationale: 'test malformed winner',
+        },
+      ],
+    }
+    await expect(
+      improve(promptProfile(), [], {
+        surface: 'memory',
+        scenarios,
+        judge: surfaceKindJudge,
+        agent: async (surface, _scenario, ctx) => {
+          ctx.cost.observe(0.0001, 'stub-agent')
+          ctx.cost.observeTokens({ input: 1, output: 1 })
+          return { text: typeof surface === 'string' ? 'text' : surface.kind }
+        },
+        memory: {
+          document: '# Durable memory\n',
+          writeBack: () => {
+            writeCalls += 1
+          },
+        },
+        generator: malformedMemory,
+        promotionGate: {
+          name: 'test-ship',
+          decide: async () => ({
+            decision: 'ship',
+            reasons: ['test candidate'],
+            contributingGates: [],
+          }),
+        },
+        budget: { generations: 1, populationSize: 1, holdoutFraction: 0.25 },
+      }),
+    ).rejects.toThrow(/winner must be text/)
+    expect(writeCalls).toBe(0)
+  })
+
+  it("surface 'memory' never writes back a held candidate", async () => {
+    let writeCalls = 0
+    const result = await improve(promptProfile(), [], {
+      surface: 'memory',
+      gate: 'none',
+      scenarios,
+      judge,
+      agent: stubAgent,
+      memory: {
+        document: '# Durable memory\n',
+        writeBack: () => {
+          writeCalls += 1
+        },
+      },
+    })
+
+    expect(result.shipped).toBe(false)
+    expect(writeCalls).toBe(0)
+  })
+
   it('a real runDir makes the loop durable: provenance lands on the filesystem', async () => {
     const { mkdtempSync, existsSync, rmSync } = await import('node:fs')
     const { tmpdir } = await import('node:os')
@@ -329,8 +474,8 @@ describe('improve() — default proposer resolution (substrate export drift guar
   })
 
   it('a surface with no zero-config default still fails loud with ConfigError', async () => {
-    // The default-proposer map covers prompt + skills only; the config surfaces
-    // (tools/mcp/hooks/code) require a caller-supplied generator. This is the
+    // Prompt, skills, memory, and rollout policy have defaults; config surfaces
+    // require a caller-supplied generator. This is the
     // designed boundary the proposer migration must NOT erase.
     const configSurfaces: ImproveSurface[] = [
       'tools',
diff --git a/src/improvement/improve.ts b/src/improvement/improve.ts
index c52bc789..7295fb23 100644
--- a/src/improvement/improve.ts
+++ b/src/improvement/improve.ts
@@ -12,6 +12,8 @@
  *
  *   - `surface: 'prompt'` → `gepaProposer` mutates `profile.prompt.systemPrompt`.
  *   - `surface: 'skills'` → `skillOptProposer` mutates a skills document string.
+ *   - `surface: 'memory'` → `memoryCurationProposer` curates a bounded durable
+ *     lesson document supplied through `opts.memory`.
  *   - `surface: 'agent-profile'` → caller-supplied proposer mutates the complete
  *     canonical AgentProfile JSON in one candidate.
  *   - `surface: 'rollout-policy'` → `rolloutPolicyProposer` mutates the
@@ -36,6 +38,7 @@
 import {
   gepaProposer,
   gitWorktreeAdapter,
+  memoryCurationProposer,
   skillOptProposer,
 } from '@tangle-network/agent-eval/campaign'
 import {
@@ -81,6 +84,7 @@ export type ImproveSurface =
   | 'subagents'
   | 'workflow'
   | 'agent-profile'
+  | 'memory'
   | 'code'
   | 'rollout-policy'
 
@@ -92,8 +96,8 @@ export type ImproveOptions<TScenario extends Scenario, TArtifact> = Omit<
    *  generator + the baseline-surface extraction shape. */
   surface?: ImproveSurface
   /** The `SurfaceProposer` that mutates the surface. When unset, the facade
-   *  picks the default for `surface` (`gepaProposer` for prompt, `skillOptProposer`
-   *  for skills); surfaces with no default REQUIRE this (fail-loud otherwise). */
+   *  picks the default for prompt, skills, memory, and rollout policy; surfaces
+   *  with no default REQUIRE this (fail-loud otherwise). */
   generator?: SurfaceProposer
   /** Gate mode. `'holdout'` (default) runs the held-out promotion gate;
    *  `'none'` is a baseline-only run (`budget.generations = 0`). */
@@ -135,6 +139,10 @@ export type ImproveOptions<TScenario extends Scenario, TArtifact> = Omit<
    *  shipped winner (the profile ref points at a file the caller owns). This is
    *  what makes skillOpt reachable through improve(). */
   skills?: ImproveSkillsOptions
+  /** MEMORY-surface wiring for a curated durable memory document. The default
+   *  deterministic proposer deduplicates and ranks lessons from findings, then
+   *  replaces its managed block instead of growing memory without bound. */
+  memory?: ImproveMemoryOptions
   /** Custom held-back-exam decision. The string `gate` above controls whether
    *  the exam runs; this callback controls how its evidence decides promotion. */
   promotionGate?: SelfImproveOptions<TScenario, TArtifact>['gate']
@@ -146,7 +154,14 @@ export interface ImproveSkillsOptions {
   /** Persist the shipped winner document (write the file the profile ref points at).
    *  Called only on a ship verdict. When omitted, the winner is still returned in
    *  `result.raw.winner.surface` for the caller to materialize. */
-  writeBack?: (winnerDocument: string) => void
+  writeBack?: (winnerDocument: string) => void | Promise<void>
+}
+
+export interface ImproveMemoryOptions {
+  /** Current durable memory text used as the measured baseline. */
+  document: string
+  /** Persist the promoted memory document. Never called on hold or error. */
+  writeBack?: (winnerDocument: string) => void | Promise<void>
 }
 
 export interface ImproveCodeOptions {
@@ -205,6 +220,8 @@ function defaultGeneratorFor(
       return gepaProposer({ llm: llmClientOptions(llm), model, target: 'agent system prompt' })
     case 'skills':
       return skillOptProposer({ llm: llmClientOptions(llm), model, target: 'agent skill document' })
+    case 'memory':
+      return memoryCurationProposer()
     case 'rollout-policy':
       // Deterministic bounded enumeration — no LLM, so `llm` is unused here.
       return rolloutPolicyProposer()
@@ -214,12 +231,13 @@ function defaultGeneratorFor(
 }
 
 /** Extract the baseline surface a driver mutates from the profile field that
- *  backs `surface`. `prompt`/`skills` are string surfaces; the config surfaces
- *  serialize the matching profile record. */
+ *  backs `surface`. Prompt, skills, and memory are text surfaces; config
+ *  surfaces serialize the matching profile record. */
 function baselineSurfaceFor(
   profile: AgentProfile,
   surface: ImproveSurface,
   skills?: ImproveSkillsOptions,
+  memory?: ImproveMemoryOptions,
 ): MutableSurface {
   switch (surface) {
     case 'prompt':
@@ -240,6 +258,11 @@ function baselineSurfaceFor(
       return JSON.stringify(profile.extensions?.[workflowExtension] ?? {})
     case 'agent-profile':
       return JSON.stringify(profile)
+    case 'memory':
+      if (!memory) {
+        throw new ConfigError("improve(): surface 'memory' requires opts.memory.document")
+      }
+      return memory.document
     case 'rollout-policy': {
       // Empty surface when the profile never opted into structural rollout: the
       // proposer reads it as "propose nothing", so the loop runs baseline-only and
@@ -265,8 +288,13 @@ function generationFailureDistiller<TScenario extends Scenario, TArtifact>(
 ): NonNullable<SelfImproveOptions<TScenario, TArtifact>['analyzeGeneration']> {
   const CAP = 12
   return async (input) => {
-    const failures: Array<{ scenario: string; composite: number; notes: string; error?: string }> =
-      []
+    const failures: Array<{
+      scenario: string
+      composite: number
+      notes: string
+      claim?: string
+      error?: string
+    }> = []
     for (const candidate of input.candidates) {
       for (const rawCell of candidate.campaign.cells) {
         const cell = rawCell as unknown as Record<string, unknown>
@@ -288,10 +316,12 @@ function generationFailureDistiller<TScenario extends Scenario, TArtifact>(
           .filter((n): n is string => typeof n === 'string' && n.length > 0)
           .join('; ')
           .slice(0, 400)
+        const claim = notes || (error ? `Scenario ${scenario} failed: ${error.slice(0, 200)}` : '')
         failures.push({
           scenario,
           composite: Number(composite.toFixed(3)),
           notes,
+          ...(claim ? { claim } : {}),
           ...(error ? { error: error.slice(0, 200) } : {}),
         })
       }
@@ -302,6 +332,19 @@ function generationFailureDistiller<TScenario extends Scenario, TArtifact>(
   }
 }
 
+/** Memory accumulates durable lessons, so keep the caller's seed findings while
+ * adding fresh judge failures. Curator proposers consume `claim`; the generic
+ * distiller retains the richer diagnostic fields for reflective proposers. */
+function memoryGenerationDistiller<TScenario extends Scenario, TArtifact>(
+  staticFindings: unknown[],
+): NonNullable<SelfImproveOptions<TScenario, TArtifact>['analyzeGeneration']> {
+  const distillFailures = generationFailureDistiller<TScenario, TArtifact>(staticFindings)
+  return async (input) => {
+    const fresh = await distillFailures(input)
+    return fresh === staticFindings ? staticFindings : [...staticFindings, ...fresh]
+  }
+}
+
 interface PreparedCodeRun {
   baseline: CodeSurface
   proposer: SurfaceProposer
@@ -424,6 +467,8 @@ function applyWinnerToProfile(
     case 'agent-profile':
       candidate = parseWinnerJson(winner, surface)
       break
+    case 'memory':
+      return profile
     case 'rollout-policy': {
       // Parse + re-validate the winner against the policy's own invariants — a
       // custom generator's malformed dial must fail loud, not persist silently.
@@ -475,6 +520,7 @@ export async function improve<TScenario extends Scenario, TArtifact>(
     rawTraceContext,
     code,
     skills,
+    memory,
     promotionGate,
     analyzeGeneration,
     ...sharedOptions
@@ -491,6 +537,9 @@ export async function improve<TScenario extends Scenario, TArtifact>(
       'improve(): the default skills optimizer requires opts.skills.document; pass the skill text or an explicit generator that understands resource refs',
     )
   }
+  if (surface === 'memory' && !memory) {
+    throw new ConfigError("improve(): surface 'memory' requires opts.memory.document")
+  }
   const usesReflectionModel = !generator && (surface === 'prompt' || surface === 'skills')
   if (usesReflectionModel) {
     assertModelAllowed(sharedOptions.llm?.model ?? defaultReflectionModel, allowedModels)
@@ -520,7 +569,8 @@ export async function improve<TScenario extends Scenario, TArtifact>(
   try {
     raw = await selfImprove<TScenario, TArtifact>({
       ...sharedOptions,
-      baselineSurface: preparedCode?.baseline ?? baselineSurfaceFor(profile, surface, skills),
+      baselineSurface:
+        preparedCode?.baseline ?? baselineSurfaceFor(profile, surface, skills, memory),
       proposer,
       budget,
       findings,
@@ -532,7 +582,9 @@ export async function improve<TScenario extends Scenario, TArtifact>(
               analyzeGeneration ??
               (rawTraceContext
                 ? rawTraceDistiller<TScenario, TArtifact>({ fallbackFindings: findings })
-                : generationFailureDistiller<TScenario, TArtifact>(findings)),
+                : surface === 'memory'
+                  ? memoryGenerationDistiller<TScenario, TArtifact>(findings)
+                  : generationFailureDistiller<TScenario, TArtifact>(findings)),
           }),
     })
   } catch (cause) {
@@ -549,18 +601,23 @@ export async function improve<TScenario extends Scenario, TArtifact>(
   }
 
   const shipped = raw.gateDecision === 'ship'
-  await preparedCode?.cleanup(shipped ? raw.winner.surface : undefined)
+  const winnerSurface = raw.winner.surface
+  await preparedCode?.cleanup(shipped ? winnerSurface : undefined)
   // When a skill DOCUMENT was optimized, the winner is document text — persist it
   // via writeBack (the profile ref points at the caller's file, unchanged) rather
   // than parsing it as a refs array. Otherwise use the standard field write-back.
-  const usedSkillDocument = surface === 'skills' && skills !== undefined
-  if (shipped && usedSkillDocument && typeof raw.winner.surface === 'string') {
-    skills?.writeBack?.(raw.winner.surface)
+  const externalDocument =
+    surface === 'skills' && skills ? skills : surface === 'memory' && memory ? memory : undefined
+  if (shipped && externalDocument) {
+    if (typeof winnerSurface !== 'string') {
+      throw new ConfigError(
+        `improve(): the shipped '${surface}' winner must be text before it can be persisted`,
+      )
+    }
+    await externalDocument.writeBack?.(winnerSurface)
   }
   const nextProfile =
-    shipped && !usedSkillDocument
-      ? applyWinnerToProfile(profile, surface, raw.winner.surface)
-      : profile
+    shipped && !externalDocument ? applyWinnerToProfile(profile, surface, winnerSurface) : profile
 
   return { profile: nextProfile, shipped, lift: raw.lift, gateDecision: raw.gateDecision, raw }
 }
diff --git a/src/improvement/index.ts b/src/improvement/index.ts
index e4547308..be6c7e7b 100644
--- a/src/improvement/index.ts
+++ b/src/improvement/index.ts
@@ -20,6 +20,7 @@ export {
 export { mcpBuildPrompt, toolBuildPrompt } from './build-prompts'
 export {
   type ImproveCodeOptions,
+  type ImproveMemoryOptions,
   type ImproveOptions,
   type ImproveResult,
   type ImproveSkillsOptions,
diff --git a/src/intelligence/improvement-cycle.ts b/src/intelligence/improvement-cycle.ts
index a53df5dd..ab9f2d21 100644
--- a/src/intelligence/improvement-cycle.ts
+++ b/src/intelligence/improvement-cycle.ts
@@ -159,8 +159,15 @@ export interface ExecuteApprovedAgentCandidateResult {
 export async function proposeAgentImprovement<TScenario extends Scenario, TArtifact>(
   options: ProposeAgentImprovementOptions<TScenario, TArtifact>,
 ): Promise<ProposeAgentImprovementResult<TScenario, TArtifact>> {
-  if (options.improvement.skills?.writeBack) {
-    throw new Error('proposeAgentImprovement cannot write a skill before human approval')
+  const writeBackSurface = options.improvement.skills?.writeBack
+    ? 'skill'
+    : options.improvement.memory?.writeBack
+      ? 'memory'
+      : null
+  if (writeBackSurface) {
+    throw new Error(
+      `proposeAgentImprovement cannot write ${writeBackSurface} before human approval`,
+    )
   }
   const analysis = await runAnalystLoop({
     ...options.analysis,
diff --git a/tests/improvement-cycle.test.ts b/tests/improvement-cycle.test.ts
index 1c24d88b..9bb705a6 100644
--- a/tests/improvement-cycle.test.ts
+++ b/tests/improvement-cycle.test.ts
@@ -151,6 +151,30 @@ afterEach(() => {
 })
 
 describe('agent improvement lifecycle', () => {
+  it('refuses memory persistence before human approval', async () => {
+    const writeBack = vi.fn()
+    await expect(
+      proposeAgentImprovement({
+        runId: 'analysis-run-memory-writeback',
+        profile: fixtureProfile(),
+        analysis: { registry: registry([finding]), inputs: {}, findingsStore: null, log: () => {} },
+        improvement: {
+          surface: 'memory',
+          memory: {
+            document: '# Durable memory\n',
+            writeBack,
+          },
+          generator: proposer,
+          scenarios,
+          judge,
+          agent,
+          budget: { generations: 1, populationSize: 1, holdoutFraction: 0.5 },
+        },
+      }),
+    ).rejects.toThrow('cannot write memory before human approval')
+    expect(writeBack).not.toHaveBeenCalled()
+  })
+
   it('binds typed candidate hooks to their equivalent measured profile commands', () => {
     expect(() =>
       assertCandidateProfileBinding(