minimem

Agent memory that's just Markdown and one SQLite file. No server, no vector database, no infrastructure.

Give any AI agent long-term memory in about 30 seconds. Your memories are plain Markdown files you own and can git — minimem indexes them into a single local SQLite file and serves hybrid semantic search to Claude Code, Codex, Cursor, or your own code over MCP.

npm install -g minimem && minimem init ~/memories

That's the whole setup. No Docker, no Postgres, no Pinecone, no daemon to babysit.

Why minimem

Zero infrastructure - The entire index is one SQLite file (sqlite-vec for vectors, FTS5 for keywords). Nothing to host, nothing to run.
Your memory, in plain files - Memories are Markdown you can read, edit, diff, and version with git. No proprietary store, no lock-in.
Local-first & portable - Commit the Markdown, .gitignore the index, and rebuild it on any machine with minimem sync.
Strong hybrid retrieval - Vector similarity fused with BM25 via reciprocal rank fusion (RRF) — the best-scoring config across the BEIR datasets we tested (results).
Bring your own embeddings - OpenAI, Gemini, or a fully local model via llama.cpp (no API key required). Guided setup picks one for you on init.
Drops into your agent - First-class MCP server for Claude Code, Codex, Cursor, and Claude Desktop, plus a Claude Code plugin — see integrations.

Benchmarks

Long-term-memory QA benchmark results for an agent memory pipeline (extract → retrieve → answer), answer model gpt-5.5:

benchmark	score	retrieval	notes
LongMemEval_S	93.0%	full pipeline	full 500-question run; ~1.9pp behind Mastra (94.87%)
LOCOMO	79.3%	minimem	10 multi-session conversations, mem0 judge
BEAM	72.7%	minimem	ICLR 2026 long-horizon memory, 500K-token conversations

The retrieval finding that earns the LOCOMO/BEAM numbers: minimem's focused hybrid retrieval beats a dump-everything KB baseline by +13pp at 500K tokens and +43pp on LOCOMO, and the gap widens with scale (the KB's context dump truncates; minimem retrieves the right notes regardless of store size). Scores are same-family-judge, not leaderboard-exact. Full methodology, the clean substrate-vs-graph decomposition, and honest per-benchmark caveats: evals/beam/RESULTS.md.

Installation

Install globally for the CLI + MCP server (recommended for most users):

npm install -g minimem

Or add it as a library dependency:

npm install minimem

Requires Node.js 22+ (uses the built-in node:sqlite).

Quick Start

# Initialize a memory directory. init detects an embedding API key in your
# environment; if it doesn't find one, it asks how you'd like to search:
#   • a hosted provider (OpenAI/Gemini),
#   • a local model (~320 MB, runs offline, no key), or
#   • keyword-only for now (upgrade anytime).
minimem init ~/memories

# Add some memories
minimem append "Decided to use PostgreSQL for the main database" --dir ~/memories

# Search your memories
minimem search "database decisions" --dir ~/memories

# Create or update a memory file
minimem upsert "memory/architecture.md" "# Architecture Notes..." --dir ~/memories

Prefer non-interactive setup? Pass --provider (e.g. minimem init ~/memories --provider local) or --yes to accept the keyword-only default in scripts/CI.

Library Usage

import { Minimem } from 'minimem';

// Create a Minimem instance
const mem = await Minimem.create({
  memoryDir: './memories',
  embedding: {
    provider: 'openai',
    openai: { apiKey: process.env.OPENAI_API_KEY }
  }
});

// Search memories
const results = await mem.search('database architecture');
for (const result of results) {
  console.log(`[${result.score}] ${result.path}:${result.startLine}`);
  console.log(result.snippet);
}

// Append to today's log
await mem.appendToday('Reviewed the API design document');

// Clean up
mem.close();

CLI Commands

Core

Command	Description
`minimem init [dir]`	Initialize a memory directory (creates files, DB, and indexes)
`minimem search <query>`	Semantic search through memories
`minimem sync`	Force re-index memory files
`minimem status`	Show index stats and provider info
`minimem append <text>`	Append to today's daily log
`minimem upsert <file> [content]`	Create or update a memory file
`minimem mcp`	Run as MCP server (stdio)
`minimem config`	View or modify configuration (`--set key=value`, `--unset key`)

Stores (cross-directory search)

Command	Description
`minimem store:add <name> <path>`	Register a store in the global manifest (`--remote <url>` for git-backed stores)
`minimem store:remove <name>`	Remove a store from the global manifest
`minimem store:list`	List all registered stores and their links
`minimem store:link <store> <target>`	Link a store to another for cross-store search
`minimem store:unlink <store> <target>`	Remove a link between stores

Git sync (multi-machine memory)

Sync memory directories through a central git repository — write on one machine, recall on another.

Command	Description
`minimem sync:init-central <path>`	Initialize a central repository for syncing
`minimem sync:init --path <name>/`	Enable sync for a memory directory
`minimem push` / `minimem pull`	Push/pull changes to/from the central repository
`minimem sync:status`	Show sync status for a directory
`minimem sync:list`	List all sync mappings
`minimem sync:remove`	Remove sync mapping for a directory
`minimem sync:conflicts`	List quarantined conflicts
`minimem sync:resolve <timestamp>`	Resolve a quarantined conflict with a merge tool
`minimem sync:cleanup`	Clean up old quarantined conflicts
`minimem sync:log`	Show sync history
`minimem sync:validate`	Validate the registry for collisions and stale mappings

Sync daemon (auto-sync)

Command	Description
`minimem daemon`	Start the sync daemon (`--background` to detach)
`minimem daemon:stop`	Stop the sync daemon
`minimem daemon:status`	Show daemon status
`minimem daemon:logs`	Show daemon logs (`--follow` to tail)

Common Options

-d, --dir <path> - Memory directory (default: current directory)
-g, --global - Use ~/.minimem as the memory directory
-p, --provider <name> - Embedding provider: openai, gemini, local, or auto

Search Options

# Search with options
minimem search "project decisions" \
  --dir ~/memories \
  --max 5 \
  --min-score 0.5 \
  --json

# Search multiple directories
minimem search "api design" \
  --dir ~/work-memories \
  --dir ~/personal-notes \
  --global

Upsert Examples

# Create/update with inline content
minimem upsert "notes.md" "# My Notes" --dir ~/memories

# Pipe content from stdin
cat document.md | minimem upsert "imported.md" --stdin --dir ~/memories

# Use heredoc for multi-line content
minimem upsert "memory/decisions.md" --stdin --dir ~/memories << 'EOF'
# Architecture Decisions

## Database
We chose PostgreSQL for its reliability and JSON support.

## API
REST with OpenAPI documentation.
EOF

Memory Directory Structure

my-memories/
├── MEMORY.md           # Main memory file (indexed)
├── memory/             # Additional memory files
│   ├── 2024-01-15.md   # Daily logs
│   ├── 2024-01-16.md
│   └── projects.md     # Topic-specific notes
├── config.json         # Configuration
├── index.db            # SQLite database with vectors (gitignored)
└── .gitignore          # Ignores index.db

Everything lives in the memory directory itself, so it's trivially portable: the Markdown files and config.json can be committed to git, while the derived index.db is ignored and rebuilt on any machine with minimem sync.

(Legacy layouts with a .minimem/ subdirectory are still read transparently.)

Configuration

Configuration is stored in config.json at the memory directory root:

{
  "embedding": {
    "provider": "auto",
    "model": "text-embedding-3-small"
  },
  "hybrid": {
    "enabled": true,
    "fusion": "rrf",
    "ftsQueryMode": "or"
  },
  "query": {
    "maxResults": 10,
    "minScore": 0.3
  }
}

The values above are the defaults. Set "fusion": "weighted" to use a score-weighted sum instead of RRF (then vectorWeight/textWeight, default 0.7/0.3, apply). A global config at ~/.minimem/config.json is layered under each directory's local config.

Environment Variables

Variable	Description
`OPENAI_API_KEY`	OpenAI API key for embeddings
`GOOGLE_API_KEY`	Google/Gemini API key for embeddings
`GEMINI_API_KEY`	Alternative Gemini API key
`MEMORY_DIR`	Default memory directory

Embedding Providers

OpenAI (recommended for quality):

export OPENAI_API_KEY=sk-...
minimem search "query" --provider openai

Gemini (good free tier):

export GOOGLE_API_KEY=...
minimem search "query" --provider gemini

Local (no API needed, requires setup):

minimem search "query" --provider local

Auto (default): Tries OpenAI → Gemini → Local based on available API keys.

MCP Server Integration

minimem can run as an MCP server for integration with Claude Desktop, Cursor, and other tools.

Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "minimem": {
      "command": "minimem",
      "args": ["mcp", "--global"],
      "env": {
        "OPENAI_API_KEY": "your-key"
      }
    }
  }
}

Multiple Memory Directories

The MCP server supports searching across multiple memory directories:

{
  "mcpServers": {
    "minimem": {
      "command": "minimem",
      "args": ["mcp", "--dir", "/path/to/work", "--dir", "/path/to/personal", "--global"],
      "env": {
        "OPENAI_API_KEY": "your-key"
      }
    }
  }
}

When multiple directories are configured:

The memory_search tool searches all directories by default
Results are merged and ranked by score
Each result shows which directory it came from
Use the optional directories parameter to filter to specific directories

Cursor

Add to Cursor's MCP settings:

{
  "mcpServers": {
    "minimem": {
      "command": "minimem",
      "args": ["mcp", "--dir", "/path/to/memories"],
      "env": {
        "OPENAI_API_KEY": "your-key"
      }
    }
  }
}

Codex (CLI & IDE)

The fastest way is the Codex CLI, which writes the config for you:

codex mcp add minimem --env OPENAI_API_KEY=your-key -- minimem mcp --global

Or add the table to ~/.codex/config.toml directly (note the snake_case mcp_servers — a typo here is silently ignored):

[mcp_servers.minimem]
command = "minimem"
args = ["mcp", "--global"]

[mcp_servers.minimem.env]
OPENAI_API_KEY = "your-key"

Verify it's live with codex mcp list. The CLI and IDE extension share this config, so you only set it up once. To scope memory to a single project, use a trusted-project .codex/config.toml and swap --global for --dir /path/to/memories.

To nudge Codex (or any agent) to actually use the memory tools, add a short block to your project's AGENTS.md:

## Memory (minimem)

Before starting a task, call `memory_search` to recall relevant prior context.
When you make a decision or learn something durable, write it to `MEMORY.md`
(or `memory/YYYY-MM-DD.md`) so it's indexed for next time.

Available MCP Tools

The MCP server exposes five tools:

Tool	Purpose
`memory_search`	Semantic search across memory files. Supports `maxResults`, `minScore`, `directories`, `detail` (`"compact"` index or `"full"` snippets), and `type` (filter by observation type: decision, bugfix, feature, discovery, context, note)
`memory_get_details`	Fetch full text for chunks returned by a compact `memory_search` — a two-phase workflow that saves tokens
`knowledge_search`	Search knowledge notes with domain/entity/confidence/type filters (see Knowledge Frontmatter)
`knowledge_graph`	Traverse knowledge-graph relationships outward from a note
`knowledge_path`	Find the shortest path between two knowledge notes via graph links

The three knowledge_* tools are additive: they only return results when knowledge-formatted notes (YAML frontmatter with type, domain, links, etc.) are present. Regular memory files work fine without them.

Typical search flow:

// 1. Compact search — lightweight index of matches
memory_search({ query: "api design decisions", maxResults: 10 })

// 2. Fetch full text only for the results that matter
memory_get_details({ results: [{ path, startLine, endLine }] })

Library API

Creating an Instance

import { Minimem, type MinimemConfig } from 'minimem';

const config: MinimemConfig = {
  memoryDir: './memories',
  embedding: {
    provider: 'openai',
    openai: {
      apiKey: process.env.OPENAI_API_KEY,
      model: 'text-embedding-3-small'  // or text-embedding-3-large
    }
  },
  hybrid: {
    enabled: true,
    fusion: 'rrf',       // default; or 'weighted' (uses vectorWeight/textWeight)
    ftsQueryMode: 'or'   // default; 'and' requires all query terms to match
  },
  watch: true,  // Auto-sync on file changes
  debug: console.log  // Optional debug logging
};

const mem = await Minimem.create(config);

Search

const results = await mem.search('database architecture', {
  maxResults: 10,
  minScore: 0.3
});

// Results include:
// - path: relative file path
// - snippet: matching text chunk
// - score: relevance score (0-1)
// - startLine, endLine: line numbers
// - heading: section heading if available

File Operations

// Append to today's daily log (memory/YYYY-MM-DD.md)
await mem.appendToday('Meeting notes: discussed API design');

// Append to specific file
await mem.appendFile('memory/project.md', 'New decision made');

// List all memory files
const files = await mem.listFiles();

// Get status
const status = await mem.status();
console.log(`Files: ${status.fileCount}, Chunks: ${status.chunkCount}`);

Sync Control

// Manual sync (usually automatic)
await mem.sync();

// Force full re-index
await mem.sync({ force: true });

Custom Embedding Providers

import { createEmbeddingProvider } from 'minimem';

// Create provider with custom settings
const provider = createEmbeddingProvider({
  provider: 'openai',
  openai: {
    apiKey: process.env.OPENAI_API_KEY,
    model: 'text-embedding-3-large',
    dimensions: 1536
  }
});

// Or use Gemini
const geminiProvider = createEmbeddingProvider({
  provider: 'gemini',
  gemini: {
    apiKey: process.env.GOOGLE_API_KEY,
    model: 'text-embedding-004'
  }
});

How It Works

Architecture Overview

┌─────────────────────────────────────────────────────────────┐
│                         minimem                              │
├─────────────────────────────────────────────────────────────┤
│  Memory Files (.md)                                          │
│  ┌─────────┐ ┌─────────┐ ┌─────────┐                        │
│  │MEMORY.md│ │daily.md │ │notes.md │  ← Plain Markdown      │
│  └────┬────┘ └────┬────┘ └────┬────┘                        │
│       │           │           │                              │
│       └───────────┴───────────┘                              │
│                   │                                          │
│                   ▼                                          │
│  ┌─────────────────────────────────────┐                    │
│  │           Chunking                   │                    │
│  │  Split by headings/paragraphs        │                    │
│  │  ~256 tokens per chunk               │                    │
│  └──────────────────┬──────────────────┘                    │
│                     │                                        │
│                     ▼                                        │
│  ┌─────────────────────────────────────┐                    │
│  │      Embedding Provider              │                    │
│  │  OpenAI / Gemini / Local             │                    │
│  │  text → [0.1, -0.3, 0.8, ...]       │                    │
│  └──────────────────┬──────────────────┘                    │
│                     │                                        │
│                     ▼                                        │
│  ┌─────────────────────────────────────┐                    │
│  │         SQLite Database              │                    │
│  │  ┌─────────┐  ┌─────────────────┐   │                    │
│  │  │  FTS5   │  │  sqlite-vec     │   │                    │
│  │  │ (text)  │  │  (vectors)      │   │                    │
│  │  └─────────┘  └─────────────────┘   │                    │
│  └──────────────────┬──────────────────┘                    │
│                     │                                        │
│                     ▼                                        │
│  ┌─────────────────────────────────────┐                    │
│  │         Hybrid Search                │                    │
│  │  Vector similarity + BM25 ranking    │                    │
│  │  Reciprocal rank fusion (RRF)        │                    │
│  └─────────────────────────────────────┘                    │
└─────────────────────────────────────────────────────────────┘

Indexing Process

File Discovery: Scans for MEMORY.md and memory/*.md files
Chunking: Splits content by Markdown headings and paragraphs (~256 tokens each)
Hashing: Each chunk gets a content hash to detect changes
Embedding: New/changed chunks are sent to the embedding provider
Storage: Chunks and vectors stored in SQLite with sqlite-vec extension
Caching: Embeddings cached by content hash to avoid re-computation

Search Process

Query Embedding: Convert search query to vector using same embedding model
Vector Search: KNN search using the sqlite-vec index (cosine distance)
Text Search: Find matching chunks using FTS5 full-text search (BM25, OR-mode)
Hybrid Fusion: Combine both ranked lists with reciprocal rank fusion (RRF)
Ranking: Sort by fused score (max-normalized), apply min-score filter

Why Hybrid Search — and Why RRF?

Pure vector search excels at semantic similarity but can miss exact matches. Pure text search finds exact terms but misses synonyms. Hybrid search combines both:

Vector: Finds conceptually related content ("database" matches "PostgreSQL")
Text (BM25): Boosts exact keyword matches ("PostgreSQL" query ranks PostgreSQL mentions higher)

The default fusion is RRF (reciprocal rank fusion): it merges the two result lists by rank position instead of mixing raw scores, so it needs no score normalization and is robust across embedding models. On BEIR retrieval benchmarks, RRF was the best (or tied-best) configuration on every dataset and embedding model tested, while score-weighted fusion collapsed when switching embedding models — see docs/RETRIEVAL-EVAL-RESULTS.md for the full numbers. A score-weighted sum is available via hybrid.fusion: "weighted".

Storage Format

SQLite database (index.db in the memory directory) with these main tables:

-- File metadata and modification tracking
CREATE TABLE files (path, source, hash, mtime, size);

-- Content chunks with embeddings and optional knowledge metadata
CREATE TABLE chunks (
  id, path, source, start_line, end_line, hash, model, text, embedding,
  type,                                      -- observation type (decision, bugfix, ...)
  knowledge_type, knowledge_id,              -- knowledge frontmatter metadata
  domains, entities, confidence,
  updated_at
);

-- Full-text search index (BM25)
CREATE VIRTUAL TABLE chunks_fts USING fts5(text, id, path, ...);

-- Vector KNN index (when sqlite-vec is available; dims are model-dependent)
CREATE VIRTUAL TABLE chunks_vec USING vec0(embedding FLOAT[<dims>]);

-- Knowledge graph edges (from knowledge frontmatter links)
CREATE TABLE knowledge_links (from_id, to_id, relation, layer, weight, ...);

Embedding Cache

Embeddings are cached by provider, model, and content hash:

CREATE TABLE embedding_cache (
  provider, model, provider_key, hash, embedding, dims, updated_at,
  PRIMARY KEY (provider, model, provider_key, hash)
);

This means:

Identical text always produces the same embedding (deterministic)
Moving/copying chunks doesn't require re-embedding
Switching files with same content is instant
The cache survives schema migrations, so re-indexing is cheap

Claude Code Plugin

A ready-to-use Claude Code plugin is included in the claude-plugin/ directory.

Quick Install

# Install minimem globally
npm install -g minimem

# Initialize global memory
minimem init --global

# Test the plugin
claude --plugin-dir /path/to/minimem/claude-plugin

Plugin Features

MCP Server: all five memory/knowledge tools (see Available MCP Tools)
Memory Skill: Auto-invoked for storing/recalling context
Commands:
- /minimem:remember <text> - Store information
- /minimem:recall <query> - Search memories
Session hooks (opt-in): inject recent memories at session start, log a marker at session end. Enable with minimem config --set hooks.sessionStart=true and hooks.sessionEnd=true.

See claude-plugin/README.md for detailed documentation.

Development

# Install dependencies
npm install

# Build
npm run build

# Run tests
npm run test:all      # All tests
npm run test          # Unit tests (vitest)
npm run test:cli      # CLI command tests
npm run test:integration  # E2E integration tests

# Development mode
npm run dev

License

MIT

Name		Name	Last commit message	Last commit date
claude-plugin		claude-plugin
docs		docs
evals		evals
media		media
scripts		scripts
src		src
.gitignore		.gitignore
AGENTS.md		AGENTS.md
CLAUDE.md		CLAUDE.md
LICENSE		LICENSE
README.md		README.md
package-lock.json		package-lock.json
package.json		package.json
renovate.json5		renovate.json5
tsconfig.json		tsconfig.json
tsup.config.ts		tsup.config.ts
vitest.config.ts		vitest.config.ts

Latest commit

History

Folders and files

Repository files navigation