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

pip install "fidelis-memory==0.2.0"
fidelis health

Or without installing, via uv:

uvx --from fidelis-memory fidelis health

Real output against a running local instance:

status: ok  |  memories: 147980  |  version: 1.0.0a2  |  calibrated: yes  |  snapshot: yes

Zero-LLM vector query:

uvx --from fidelis-memory fidelis query "test query" --limit 2

2 memories:

  [1]  score 0.686
      User is 'test'

  [2]  score 0.686
      [fact] Reliable testing procedure for Google Rich Results Test: navigate fresh to page, fill test URL textbox, press Escape, click test URL button, wait ~20s for results

Output shape

• health: one-line status, memory count, version, calibration/snapshot state
• query / recall / recall-hybrid: ranked {text, score} memories, plus a method field naming the retrieval path taken (filter, fallback_*, decompose_N[_v])
• HTTP endpoints mirror the CLI 1:1 (/health, /recall, /recall_hybrid, /query, /store, /add, /snapshot, /replay)

Common gotchas

• Executable name is fidelis, not fidelis-memory — uvx fidelis-memory fails; use uvx --from fidelis-memory fidelis <cmd>.
• The PyPI project literally named fidelis is unrelated; install fidelis-memory.
• The optional filter/extraction tiers call out to Ollama or an Anthropic/OpenAI-compatible endpoint — the default /recall_b and /query paths do not.

More

Full docs, HTTP reference, and module map: https://github.com/hermes-labs-ai/fidelis