Skip to content

Releases: jaylfc/taosmd

v0.3.0

09 Jun 15:55
@jaylfc jaylfc
v0.3.0
60a8d5a

Choose a tag to compare

View all tags
v0.3.0 Latest
Latest

First PyPI release. pip install taosmd.

Highlights since v0.2.0

Recipes engine — tier-aware recommend(device_info) grounded in the tri-judge benchmarks, a Recipe dataclass with JSON-Schema export, per-agent applied recipe + retrieval config (write-through), a global default_recipe setting, and search() that resolves and applies the active recipe with a non-blocking bge-v2-m3 reranker download that degrades gracefully when absent. SP4 recipe methods on the MemoryBackend ABC.

Project identity — git-remote project fingerprinting, project-scoped storage with opt-in cross-agent search, librarian discovery (list_projects/list_shelves), wired across the Python API, HTTP, MCP, CLI, and remote client, plus a dashboard Projects view.

A2A bus — realtime wake via a2a-watch (SSE stream) and a2a-bridge (exec-on-message), an all-channels mode for both, and an atomic poll state-file write with graceful bus-unreachable handling.

Retrieval — fusion and candidate-pool size threaded to the vector source.

Full benchmarks and methodology in docs/benchmarks.md.

Assets 2
Loading

taOSmd v0.2.0

08 Jun 04:04
@jaylfc jaylfc
v0.2.0
7bf79b0

Choose a tag to compare

View all tags
taOSmd v0.2.0

First tagged release of taOSmd, a framework-agnostic, local-first AI memory system that runs offline on modest hardware (8 GB+ RAM, Python 3.10+).

Activation surfaces (available today)

  • Python API and CLI (taosmd, taosmd serve, taosmd mcp, taosmd reconcile, taosmd install-skill).
  • Local HTTP/REST API via taosmd serve (search, ingest, health, and the A2A bus endpoints).
  • MCP server over stdio via taosmd mcp (memory tools plus the A2A bus tools); the MCP SDK is an optional [mcp] extra.
  • Web dashboard served offline by taosmd serve for search, pending review, and health.

Core

  • Append-only, zero-loss archive: the verbatim turn is stored first; summaries and structure are layered on top, never over the source.
  • Vector memory (local ONNX embeddings, hybrid keyword plus vector search, optional binary-quantized vectors).
  • Temporal knowledge graph and a librarian retrieval layer.
  • Per-agent isolation (each agent gets its own shelf) with cross-agent reads.
  • Correction and supersede across both the knowledge graph and the vector layer, while the raw archive row is retained.
  • Optional remote client mode: point a local CLI at a shared taosmd serve instance over your own network.

Install

Install from source (verified on a clean environment):

git clone https://github.com/jaylfc/taosmd.git
cd taosmd
pip install -e .

A PyPI release is planned. The one-line bootstrap in the README additionally installs Ollama and the embedding and LLM models and is still being validated across clean machines.

Benchmarks

Maintainer-published benchmark notes and methodology are in docs/benchmarks.md. These are maintainer-run results on a local low-end reference stack, not independent third-party validation.

License

MIT.

Loading