MCP & AI clients

The memory has no search UI to learn. It speaks the Model Context Protocol over stdio, so the AI tools you already use read it directly.

Connect a client

Everything goes through one launcher, bin/lavox-mcp.sh. It sets up its own slim Python environment on first run, so there is no manual setup beyond cloning the repo.

Claude Code as a plugin

/plugin marketplace add lavox-app/lavox
/plugin install lavox-memory@lavox

The plugin registers the server and adds three commands: /lavox-memory:memory-search, :decisions and :remember.

Claude Code with plain MCP

claude mcp add lavox-memory -- /path/to/lavox/bin/lavox-mcp.sh

Codex CLI

# ~/.codex/config.toml
[mcp_servers.lavox-memory]
command = "/path/to/lavox/bin/lavox-mcp.sh"

Cursor

// .cursor/mcp.json
{ "mcpServers": { "lavox-memory": { "command": "/path/to/lavox/bin/lavox-mcp.sh" } } }

The seven tools

Tool	What it does
`search`	Fusion search over the whole memory; the workhorse. Supports filtering by kind and time, and can include superseded history on request.
`fetch`	The full context of one chunk: what was said before and after. Use it when a snippet is relevant but truncated, or to quote exactly.
`timeline`	What happened in a time range: browsing, not searching.
`stats`	Corpus counts: recordings, chunks, assertions.
`remember`	Explicitly save a fact or decision. Warns when a similar memory already exists.
`correct`	Supersede an assertion with a corrected one. Never deletes; history stays queryable.
`profile`	A condensed profile of the user: stable facts and preferences.

The two core tools are named search and fetch deliberately: remote MCP connectors for ChatGPT expect exactly these names, so the planned hosted connector will work without renaming anything.