Every AI session starts blank. Decisions, architecture choices, failed approaches — gone when you close the terminal. mnemo watches your sessions, extracts what matters, and injects it back into Claude Code at session start. More tools coming soon.

Why mnemo?

• Zero-config — npx mnemo-ai init and you're done. No YAML, no dashboard, no setup wizard.
• TypeScript-native — built in TypeScript, for TypeScript developers. First-class MCP support.
• Dev-tool-specific — purpose-built for AI coding workflows. Not a generic note-taking app.
• Auto-capture — memories are extracted automatically from your sessions. You don't write anything manually.

Supported tools

| Tool | Status | |------|--------| | Claude Code | Supported | | Cursor | Supported | | Windsurf | Coming soon | | Codex CLI | Supported | | Aider | Coming soon | | Continue | Coming soon | | Cline / Roo Code | Coming soon |

Install

npm install -g mnemo-ai

Requires Node.js 20+ and an Anthropic API key (get one free).

Setup

# 1. Set your API key
export ANTHROPIC_API_KEY=sk-ant-...

# Add it to your shell profile so it persists:
echo 'export ANTHROPIC_API_KEY=sk-ant-...' >> ~/.zshrc

Quick start

cd your-project

# 1. Set up mnemo + MCP for your AI tools
mnemo init

# 2. Start watching (run alongside your AI coding session)
mnemo watch

# 3. Later — ask mnemo what it remembers
mnemo recall "how did we handle auth?"
mnemo recall "why did we choose postgres?"

# 4. See everything captured
mnemo recall

# 5. Check your memory stats
mnemo status

How it works

mnemo runs a background daemon that watches your AI coding session files. It captures memories on two triggers: 30 seconds of inactivity (between pauses) and a 5-minute safety net (during long continuous sessions). Extracted memories include:

| Type | What it captures | |------|-----------------| | decision | Architecture/implementation choices and why | | rejection | Approaches tried and abandoned | | pattern | Recurring conventions for this codebase | | reference | Where important things live | | error | Error patterns and their fixes |

Memories are stored locally in ~/.mnemo/memories.db — your data never leaves your machine (free tier).

Data flow

AI session files (.jsonl)
  → mnemo watches for changes (chokidar)
  → flushes on 30s inactivity OR every 5 minutes
  → extracts decisions, patterns, references (AI-powered)
  → stores in local SQLite with FTS5 + semantic embeddings
  → next session: MCP injects memories automatically

Commands

mnemo init              # set up MCP integration for your AI tools
mnemo watch             # start session watcher daemon
mnemo recall            # show recent memories for this project
mnemo recall <query>    # semantic search + AI-synthesized answer
mnemo status            # memory stats for current project
mnemo forget <id>       # delete a specific memory
mnemo mcp               # start MCP server (used internally by AI tools)
mnemo prompt            # shell prompt indicator (for PS1 integration)

Status Awareness

mnemo keeps you informed about its state — both in your terminal and inside your AI tools.

Shell prompt indicator

Add this to your ~/.zshrc or ~/.bashrc:

export PS1='$(mnemo prompt)'$PS1

When the watcher is active, your prompt shows [mnemo]:

[mnemo] ~/my-project (main) $

MCP status line

Every time your AI tool loads context, mnemo reports its state:

[mnemo] watching | 42 memories | branch: main | last: 2m ago

Your AI knows whether mnemo is active, how many memories exist, and when the last capture happened.

MCP Integration

After mnemo init, Claude Code automatically loads memories at session start via MCP. Your AI arrives knowing:

• Key architectural decisions
• What's been tried and failed
• Patterns and conventions
• Where things live

MCP Tools

| Tool | Description | |------|-------------| | mnemo_context | Load project memories + status line at session start | | mnemo_save | Save a new decision, pattern, or reference |

Memory types

[DECISION]  ★8  Using Postgres over MySQL for JSONB support
            → Need flexible schema for user prefs without migrations

[REJECTION] ★7  Tried Prisma migrations for this schema, abandoned
            → Too rigid for our multi-tenant approach, switched to raw SQL

[PATTERN]   ★6  Always validate with Zod before hitting the DB layer
            → files: src/lib/validators.ts

[REFERENCE] ★5  Auth logic lives in /lib/auth.ts, not /api/auth

Importance scale: 1 (minor detail) → 10 (critical architecture decision)

Environment variables

| Variable | Required | Description | |----------|----------|-------------| | ANTHROPIC_API_KEY | Yes | Your Anthropic API key | | MNEMO_EXTRACT_MODEL | No | Model for extraction (default: claude-sonnet-4-5) | | MNEMO_RECALL_MODEL | No | Model for recall synthesis (default: claude-sonnet-4-5) | | MNEMO_EMBED_MODEL | No | Model for embeddings (default: claude-haiku-4-5) |

Typical session cost: < $0.01

Project structure

src/
  index.ts              # CLI entry point
  types.ts              # shared type definitions
  commands/
    init.ts             # mnemo init — MCP setup
    watch.ts            # mnemo watch — session watcher
    recall.ts           # mnemo recall — memory search
    status.ts           # mnemo status — stats
    forget.ts           # mnemo forget — delete memory
  core/
    extractor.ts        # AI-powered memory extraction
    watcher.ts          # session file watcher daemon
    embeddings.ts       # semantic embeddings for search
    git.ts              # git context (branch, project name)
  db/
    store.ts            # SQLite + FTS5 storage layer
  mcp/
    server.ts           # MCP server for AI tool integration
  utils/
    apikey.ts           # API key validation

Pricing

| Tier | Price | What you get | |------|-------|-------------| | Solo | Free | Unlimited local memory, MCP, all CLI commands, FTS5 + semantic search, git branch-aware scoping | | Pro | $9/dev/mo | Cloud sync, team memory vault, memory REST API, priority extraction | | Team | $29/5 devs/mo | Shared project vault, admin controls, GitHub/GitLab org integration |

Built with

• Node.js + TypeScript — CLI and MCP server
• SQLite + FTS5 — local storage, zero config, full-text search
• Anthropic API — memory extraction + recall synthesis
• MCP SDK — AI tool integration (works with any MCP-compatible tool)
• chokidar — real-time session file watching

License

MIT

Made by GE Labs LLC