nexor-mneme 0.1.28

Open memory engine + code intelligence for AI agents

nexor-mneme

Persistent memory + code-intelligence for AI agents — a knowledge layer you self-host.

nexor-mneme is the open memory engine of the Nexor project (mneme, μνήμη — "memory"). It gives agents long-term semantic memory and a queryable structural map of your codebase, so they answer "what did we decide" and "what calls this function / what's the blast radius" without re-reading everything.

Dependency-light core; heavier capabilities are optional extras. Run it standalone — no dependency on anyone's personal orchestrator or daemon.

What's inside

• Semantic memory — hybrid (dense+sparse) retrieval over Qdrant, cross-encoder reranking, forgetting, contradiction & entity resolution, auto-profile.
• Progressive layering (L0→L3) — a semantic pyramid with provenance + deterministic drill-down from an abstraction to its evidence.
• Code structure graph — multi-language (Python via stdlib ast; JS/TS/Go via tree-sitter), precise resolution via SCIP ingest, blast-radius queries, PageRank repo-map, community detection, incremental cache — landed in Apache AGE.
• Symbolic task memory — Mermaid task canvases with offload-by-reference.
• Obsidian export — a human-browsable, interlinked vault of the whole graph.
• Standalone service — a FastAPI server + docker-compose (server extra).

Install

pip install nexor-mneme                 # core memory + code graph (Python)
pip install "nexor-mneme[codegraph]"    # + tree-sitter (JS/TS/Go)
pip install "nexor-mneme[scip]"         # + SCIP precise resolution
pip install "nexor-mneme[server]"       # + FastAPI standalone service

Run your own instance

git clone https://github.com/nexorhq/nexor-mneme && cd nexor-mneme/deploy
docker compose up -d                    # API on http://localhost:8200
curl http://localhost:8200/health

See the deploy guide.

License

MIT — see LICENSE.