fidelislisted

# fidelis fidelis is local-first retrieval memory for AI agents and Claude Code. The default zero-LLM path returns stored passages verbatim — no LLM call, no rephrasing risk. It layers a compressed markdown snapshot index (~741 tokens) on top of two-stage recall (BM25 + dense + reciprocal-rank-fusion) for cross-reference queries flat memory or vector-only RAG miss. ## Use it for - Standing up a local HTTP memory server an agent host queries before/instead of a model call - Zero-LLM recall (`recall-hybrid`, `recall_b`) where verbatim, unmodified text must come back - Bulk-seeding a corpus from markdown/text files and querying it immediately - Wiring an MCP memory tool into Claude Code, Codex, GitHub Copilot CLI, Gemini CLI, or OpenClaw via `fidelis mcp install` ## Do not use it for - A hosted, multi-tenant memory platform (this is a local process/service) - A guarantee that retrieval accuracy transfers unchanged to a different corpus or workload — published numbers are project measurements on LongMemEval-S, not independent replication - Proving a model's stored claim is factually true — fidelis returns what was stored, it does not fact-check it ## Quickstart ```bash pip install "fidelis-memory==0.1.0" fidelis health ``` Or without installing, via [uv](https://docs.astral.sh/uv/): ```bash uvx --from fidelis-memory fidelis health ``` Real output against a running local instance: ``` status: ok | memories: 147980 | version: 1.0.0a2 | calibrated: yes