v2.1.2 · MCP server · local-first

Persistent memory for AI coding agents.

A local-first MCP server that gives Claude Code, Codex, Cursor and any stdio-MCP client durable project memory — decisions, patterns, pitfalls, architecture notes, and team-shared knowledge.

Install View on GitHub

~ — bash

$ npm install -g @lfrmonteiro99/memento-memory-mcp
$ memento-mcp install
$ memento-mcp import claude-md
$ memento-mcp ui  # open the local inspector

Claude Code
Codex
Cursor
Any stdio-MCP client

// the problem

Agents are powerful. They are also amnesiac.

They forget why.

The decision you defended in three meetings? Gone next session.

You repeat context.

Every chat starts with the same paragraph of project lore.

Tokens drain.

Hand-pasted memory burns context windows you'd rather spend on code.

// what it does

A typed memory layer that survives sessions.

Six memory types, ranked and retrieved by intent. Your agent stops guessing and starts remembering.

factdecisionpreferencepatternarchitecturepitfall

Local-first by default

SQLite + FTS5 search. No vector database required. No mandatory cloud account. Your project history doesn't quietly leak to a SaaS.

Team memory through git

Team-scoped memories serialize to .memento/memories/. Review, diff, commit, share — like normal project files.

Token-aware injection

Adaptive ranking with utility feedback. The right context lands in the prompt — not all of it, the right of it.

Optional embeddings

FTS5 and vector results merged through one ranker. Bring your own key when you want semantic search. Skip it when you don't.

Privacy that holds

<private>…</private> regions are excluded from search, embeddings, sync and LLM calls. Secret scrubbing at write time.

Per-project policy

.memento/policy.toml controls required tags, banned patterns, retention, and vault promotion — checked into the repo.

Session summaries

Capture the end-of-session signal deterministically, or with optional LLM-assisted summaries via Anthropic or OpenAI.

Local web inspector

memento-mcp ui opens a browser inspector for memories, sessions, projects, sync drift, analytics and health.

// how it works

Capture. Store. Rank. Inject.

Agent activity

Decisions, pitfalls and patterns surface during a session.

Typed store

Memories are written to local SQLite with type, scope and tags.

Ranked retrieval

FTS5 + optional embeddings, scored by type, recency and utility.

Context injection

The right slice lands in your agent's prompt at the right time.

decision.json pitfall.json

{
  "type": "decision",
  "title": "Use repository classes for complex SQL",
  "body": "Keeps business logic out of controllers; easier perf tuning.",
  "scope": "project",
  "tags": ["architecture", "backend"],
  "created_at": "2026-04-25T12:04:11Z"
}

// local-first

Your memory. Your machine. Your repo.

Local SQLite. No required cloud.
FTS5 search out of the box.
Optional embeddings — opt-in, your key.
Team sync via git, not a SaaS.
Private regions excluded from sync, search and LLM calls.
Secret scrubbing at write time.

// 60-second tour

From npm to remembering, in four steps.

Install

npm install -g @lfrmonteiro99/memento-memory-mcp

Wire your MCP client
```
memento-mcp install
```
Configures Claude Code, Codex, Cursor, or any other stdio-MCP client.

Import existing project memory

memento-mcp import claude-md --dry-run
memento-mcp import claude-md --no-confirm

Inspect & share

memento-mcp ui              # local browser inspector
memento-mcp sync init       # share team memory via git
memento-mcp sync pull

Get it on GitHub View on npm

" Stop pasting the same project explanation into every new chat like a medieval scribe with npm installed. "

— the README, unrepentantly