# Fidelis Memory [Health: Active]

**Category:** 🧠 Knowledge & Memory  
**Repository:** https://github.com/hermes-labs-ai/fidelis  
**GitHub Stars:** 24  
**Views:** 1  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/fidelis-memory

## Description
Recall locally stored agent memories through an MCP-compatible host.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "fidelis-memory": {
    "command": "uvx",
    "args": ["--from","fidelis-memory","fidelis-mcp"]
  }
}
```

## Documentation

## What the Fidelis Memory MCP server does

Fidelis Memory MCP server gives compatible agents access to notes stored on the local machine. It can save Markdown and plain-text content, retrieve relevant passages, apply corrections that supersede earlier records, fetch stored material, and inspect memory state. The verbatim write and retrieval paths are intended for cases where exact wording, conditions, or warnings must remain available to later sessions.

The default MCP recall path uses vector search. An explicit hybrid mode combines keyword and vector results. Its `zero_llm` tier does not use a generative model to rewrite the retrieved text, although local embeddings still require Ollama. Optional extraction and model-assisted routes may transform or filter inputs, so they are less suitable when preservation of the supplied wording is the priority.

## How it works

`fidelis init` starts a per-user background service: launchd on macOS or a user systemd service on Linux. The service runs `fidelis-server` at login and stores data in `~/.cogito/store` by default. A watcher sends files to that service for ingestion; it also records processed file paths in `~/.fidelis/watched.json`.

After a note is available locally, an MCP client can call Fidelis tools when it determines that memory is relevant. Installing the connection does not force recall on every turn. Retrieved content is still subject to being outdated or incorrect, and the model provider handling the agent’s response may receive that content.

## Setup and configuration

The documented quick start supports Python 3.10 or newer on macOS or Ubuntu. Create a virtual environment, install `fidelis-memory[hybrid]`, run `fidelis init`, and make Ollama available with the `nomic-embed-text` model. Ingest a notes directory with `fidelis watch <directory> --once`, then test retrieval with `fidelis recall-hybrid`.

Connect an agent with `fidelis mcp install --client codex`, or use `fidelis mcp install` for Claude Code. Restart the client before using the memory tools. Keep the virtual environment after initialization because the background service uses it. `fidelis health` can help diagnose retrieval problems, including a missing Ollama embedding model.

Uninstalling the macOS service removes its launchd configuration but not stored data. On Linux, disable and remove the installed user unit before reloading systemd. Deleting `~/.cogito/store` removes all Fidelis data, while deleting the watched-file record only removes ingestion history.

## Tools and capabilities

The MCP connection exposes tools for:

- Recalling relevant stored memories.
- Storing new memory content.
- Correcting an existing memory while retaining its history.
- Fetching stored content.
- Inspecting memory and validity information.

Fidelis Memory MCP server is suited to development notes, operational conditions, decisions, and other local context that agents may need in later sessions. The project also documents connections for Cursor and Gemini CLI, in addition to Codex and Claude Code.

## Limitations and notes

The 0.3.0rc1 release is a single-machine pre-release, not a hosted team memory service. Ollama and a local embedding model are required for the documented retrieval setup; no model API key is required in that path. Retrieval benchmarks described by the project measure whether material was found, not whether an agent’s final answer was accurate.

The store is shared across everything saved through Fidelis, so removing it is a destructive action. Notes can become stale, and optional model-assisted processing can change or filter input. Use a verbatim path when preserving the original passage matters.

_Full upstream README: https://allmcps.com/mcp/fidelis-memory/readme_