# Fidelis Memory
Local agent memory with zero-LLM retrieval by default, verbatim records, and correction history.

Version: 0.3.0rc1 (pre-release). Distribution: fidelis-memory. Import/CLI: fidelis.
Repository: https://github.com/hermes-labs-ai/fidelis

Install on macOS or Linux with Python 3.10+, Ollama, and nomic-embed-text:

```bash
python3 -m pip install "fidelis-memory==0.3.0rc1"
fidelis init
fidelis mcp install --client codex
```

Other installers: Claude Code (default), copilot, gemini, openclaw.
MCP tools: fidelis_recall, fidelis_store, fidelis_correct, fidelis_get,
fidelis_recent, fidelis_health. Old fidelis_query/orient and cogito_* MCP names
are retired in this release. Restart clients after upgrading.

Default MCP recall calls HTTP /query with local learned embeddings and no
generative LLM. mode="thorough" calls /recall_hybrid with tier="zero_llm".
Corrections preserve old records with explicit temporal state; as_of queries
apply validity dates. Read acknowledgements: stored, duplicate, rejected,
and queued are distinct outcomes. Recall is evidence, not truth verification.

Historical benchmark scores do not measure the redesigned default path.
Full LLM/QA evaluation is post-release work; consult current README evidence.

Service: http://127.0.0.1:19420. FIDELIS_PORT configures client and server.
Store: ~/.cogito/store. COGITO_STORE_PATH and FIDELIS_QUEUE_DIR redirect state.
Ollama is required for memory operations, but health and MCP discovery remain
available without it. Local storage is not application-encrypted; screening
cannot detect every secret. Not a network-authenticated multi-tenant service.

Reference: docs/full-reference.md
Release notes: docs/releases/0.3.0rc1.md
Security: SECURITY.md
