Install local memory and verify one retrieval.

Install Ollama and its local embedding model first. Then install the released package, initialize the local service, ingest a disposable note, and query it. fidelis init installs the persistent fidelis-server service under launchd or systemd; it starts on boot, logs to ~/.fidelis/server.log, and stores Chroma and SQLite data in ~/.cogito/.

$ ollama pull nomic-embed-text
$ python3 -m pip install "fidelis-memory==0.2.0"
$ fidelis init
$ fidelis health

$ mkdir -p /tmp/fidelis-verify
$ printf '%s\n' 'Release decision: use the blue deployment.' > /tmp/fidelis-verify/note.md
$ fidelis watch /tmp/fidelis-verify --once
$ fidelis query 'Which deployment did we choose?'

Decision boundary. Windows is not supported by this release. The setup needs a reachable local Ollama service; it does not make an agent call recall automatically. Stop the persistent service with fidelis init --uninstall.

After the local retrieval works, run the command for one client below. These commands are available in the tagged 0.2.0 package. Restart that client and inspect its MCP tools; a registered server alone does not prove that the local service can retrieve your note.

$ fidelis mcp install

$ fidelis mcp install --client codex

$ fidelis mcp install --client copilot

$ fidelis mcp install --client gemini

$ fidelis mcp install --client openclaw

Cursor source preview. Cursor support is on the repository's current main, which uses a 0.3.0rc1 prerelease; it is not in the tagged 0.2.0 setup above. If a 0.2.0 service is already running, stop it with that installation's fidelis init --uninstall before switching. Run python3 -m venv ~/.venvs/fidelis-cursor-preview and source ~/.venvs/fidelis-cursor-preview/bin/activate, then python3 -m pip install git+https://github.com/hermes-labs-ai/fidelis.git@main, fidelis init, and fidelis health so the local service matches that source. Register Cursor with fidelis mcp install --client cursor. It writes the native ~/.cursor/mcp.json entry, preserving other servers. Restart Cursor, open Customize → MCP, and confirm six Fidelis tools. The same source has a portable Agent Plugin for clients that load plugin.json and mcp.json; use one Cursor route per profile. Read the current source and prerequisites.

Ask for fidelis_health, then fidelis_recall for a distinctive phrase in the note you ingested. The returned passage should preserve the note's wording. The agent chooses when to call MCP tools; registration does not make recall automatic. Remove a client entry with fidelis mcp uninstall --client CLIENT.

After the local service can retrieve your note, Pi 0.87.1 or newer (Node.js 22.19 or newer) can load the repository's native extension. This source path is separate from the tagged 0.2.0 package setup above.

$ pi install git:github.com/hermes-labs-ai/fidelis@main
$ pi list

Restart Pi or run /reload. Ask about a distinctive note you already ingested and confirm its exact text appears in a fidelis-pi-recall message before the model answers. pi list confirms registration; it does not prove recall worked.

Boundary. The extension requests local /recall_b context before user turns with at least three non-whitespace characters. It does not add the six MCP tools, write memory, or create another memory store. An unavailable service warns and lets the turn continue, while an empty result adds no context. Temporal status is unchecked, so inspect the source and current validity of a retrieved note before relying on it. Remove the extension with pi remove git:github.com/hermes-labs-ai/fidelis@main. Read the merged Pi setup and limits.

Use fidelis health before debugging client registration. If it cannot report status: ok, make sure Ollama is running and the embedding model has been pulled, then rerun fidelis init. For a missing result, ingest a known file with watch --once and query a distinctive phrase from it.

The released service is local-first and the default retrieval path does not call an LLM. It does not turn stored content into verified facts, provide Windows support, or establish a compliance posture for a deployment. Read the full reference and user-fit guide before using it for a shared or regulated memory store.

Open-source catalog · Continuity and memory · Product overview