Mimir

The Problem

You spin up three Claude Code agents on a refactoring task. Agent A discovers the API uses snake_case internally. Agent B discovers the same thing — the hard way — and introduces a bug. Agent C? It breaks the same thing all over again.

Your agents are amnesiacs. Every session starts from zero. No shared context, no institutional memory, no coordination. You become the bottleneck — manually relaying context, copying error messages between terminals, babysitting each agent so they don't repeat each other's mistakes.

Mimir fixes this. When Agent A learns something, Agent B and C know it before they even start.

How It Works

Agent A starts  ──→  hook  ──→  mimir daemon  ──→  DuckDB
Agent A marks "API uses snake_case"  ──→  saved

Agent B starts  ──→  hook  ──→  mimir daemon  ──→  reads A's marks
                                     └──→  injects as additionalContext
                                           "⚠️ API uses snake_case"

Agent C starts  ──→  same thing. Zero repeated mistakes.

Mimir hooks into Claude Code's lifecycle events. Every time an agent starts, stops, reads a file, or edits code — Mimir captures the context. When the next agent spins up, it receives everything relevant automatically via additionalContext injection.

No extra prompting. No copy-paste. It just works.

Quick Start

# Install globally
npm install -g mimir

# Initialize in your project (installs hooks, starts daemon)
mimir init .

# Restart your Claude Code session — that's it

Three commands. Your agents now share a brain.

Prerequisites: Node.js 22+, jq — brew install jq (macOS) / sudo apt install jq (Ubuntu/Debian)

Upgrading

# If installed globally via npm
npm update -g mimir

# If installed from source
git pull && pnpm install && pnpm build

The database schema auto-migrates on daemon startup — no manual migration needed. After upgrading, restart the daemon:

mimir stop && mimir start

Features

Self-Marking Knowledge System

Agents automatically mark important discoveries during work — warnings, decisions, and learnings. These marks are stored in DuckDB and surfaced to future agents working on related files.

Agent discovers: "DuckDB COUNT(*) returns BigInt — wrap with Number()"
    → saved as warning
    → every future agent touching DuckDB gets this injected automatically

RAG-Based Smart Context

Not just keyword matching. Mimir uses Cloudflare Workers AI (bge-m3, 1024-dim embeddings) + DuckDB vector similarity to find semantically relevant past knowledge.

• Push: Top 5 relevant marks auto-injected when agents start
• Pull: Agents search past marks on-demand via MCP tools
• Fallback: Graceful ILIKE text search when embeddings aren't available

Swarm Mode

Launch multi-agent tmux sessions with built-in coordination:

mimir swarm -a "backend:sonnet,frontend:sonnet" -t "Refactor auth module"

Each agent gets its own tmux pane, a dedicated messaging channel, and automatic context sharing. The orchestrator stays lean — it only sees final reports.

Agent Teams Compatible

Works seamlessly with Claude Code's experimental Agent Teams. Teammates fire the same hooks — zero code changes needed.

{
  "env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" }
}

MCP-Powered Messaging

Inter-agent communication through the Model Context Protocol. Agents can:

• Send/receive direct messages
• Search past observations from any session
• Mark discoveries for future agents
• Track promotion candidates for permanent rules

VSCode / Cursor Extension

A companion extension with sidebar navigation, real-time agent monitoring, orchestration controls, and a full web dashboard — all embedded in your editor.

Knowledge Lifecycle

Marks have a lifecycle. Active marks get pushed to new agents. Resolved marks are excluded from injection but remain searchable. Frequently recurring patterns get promoted to permanent rules files.

Hot   → Current session marks (auto-injected)
Warm  → Past session marks (RAG search + injection)
Cold  → Agent MEMORY.md files (persistent patterns)
Perm  → .claude/rules/ (promoted, always loaded)

Curator Agent

Automated knowledge hygiene. Run mimir curate to audit marks, cross-pollinate learnings between agents, and promote recurring patterns to rules.

mimir curate              # Interactive
mimir curate --background # Background (tmux)
# Or schedule: 0 */6 * * * mimir curate --background

Architecture

Claude Code Session
├── Hook events ──→ hook.sh (jq + curl, 3s timeout)
│                      └──→ mimir daemon (Hono, port 3100)
│                              ├── DuckDB (embedded, WAL mode)
│                              ├── RAG embeddings (Cloudflare bge-m3)
│                              ├── WebSocket (real-time UI updates)
│                              └── MCP server (agent messaging + marks)
├── Web UI (React 19 + Vite + TailwindCSS 4)
│   └── Dashboard, Agents, Tasks, Marks, Skills, Curation
└── VSCode Extension (sidebar + webview panels)

Tech Stack

CLI Reference

mimir start                  # Start the daemon (background)
mimir stop                   # Stop the daemon
mimir status                 # Show active sessions and agents
mimir init [path]            # Install hooks + templates + register project
mimir ui                     # Open web dashboard in browser
mimir logs [-n 50] [-f]      # Tail daemon logs
mimir curate [--background]  # Run knowledge curation agent
mimir swarm -a "a:opus,b:sonnet" [-t "task"]  # Launch multi-agent tmux session
mimir mcp                    # Run MCP server (stdio mode)

Configuration

Environment Variables

cp .env.example .env

RAG embeddings are optional. Without them, Mimir falls back to text-based ILIKE search — still works, just less semantically aware. See .env.example for details.

What mimir init Sets Up

• .claude/settings.local.json — Hook configuration + Agent Teams
• .mcp.json — MCP server for inter-agent messaging
• .claude/skills/ — Self-marking, self-search, self-memory, and more
• .claude/agents/ — Curator agent definition
• .claude/rules/ — Team coordination rules

Troubleshooting

Daemon won't start / port already in use

• Check if another mimir instance is running: lsof -ti :3100
• Kill the orphan process: kill $(lsof -ti :3100)
• Or use a different port: mimir start --port 3200

Hooks not firing after mimir init

• Restart your Claude Code session (hooks are loaded at session start)
• Verify hooks are installed: cat .claude/hooks.json
• Check daemon is running: mimir status

DuckDB corruption / WAL recovery

• mimir auto-recovers on startup by removing stale WAL files
• If the daemon won't start, try: mimir stop && mimir start
• As a last resort, delete data/mimir.duckdb.wal and restart

Agents appear stuck / zombie agents

• Agents active > 2 hours are automatically cleaned up
• Manual cleanup: use the Kill button in the Web UI (Agents page)
• Or restart the daemon: mimir stop && mimir start

RAG/embeddings not working

• Embeddings are optional — mimir falls back to text search (ILIKE)
• To enable RAG: set CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_TOKEN in your environment
• See .env.example for details

Web UI shows blank page

• Ensure the daemon is running: mimir status
• Check logs for errors: mimir logs

Database growing too large

• Check size: mimir status (shows DB size)
• Clean old data: mimir prune --older-than 30
• Use --dry-run to preview: mimir prune --older-than 30 --dry-run

Roadmap

• Hook-based lifecycle tracking (session, agent, file changes)
• Self-marking knowledge system (warnings, decisions, discoveries)
• RAG-based smart context injection (Cloudflare bge-m3 + DuckDB vss)
• Mark lifecycle (active/resolved + promotion to rules)
• MCP server for agent messaging and mark search
• Swarm mode (multi-agent tmux orchestration)
• VSCode/Cursor extension
• Curator agent for automated knowledge hygiene
• npm publish + npx mimir init . one-liner setup
• Multi-project knowledge sharing
• Conflict-aware mark merging for parallel agents
• Slack integration for async team notifications
• Dashboard analytics (agent efficiency, mark hit rates)
• Plugin system for custom hook handlers
• Cloud-hosted daemon option
• First-class Windows support

Contributing

PRs are always welcome. Here's the quick workflow:

git clone https://github.com/SierraDevsec/mimir.git
cd mimir
pnpm install
pnpm dev          # Start dev server
pnpm test         # Run tests
pnpm build        # Full build (TypeScript + Vite)

Project Structure

src/
  cli/          — CLI commands (commander.js)
  hooks/        — Hook script (bash + jq + curl)
  server/       — Hono server, routes, services, DuckDB
  mcp/          — MCP server (messaging + marks)
  web/          — React SPA (dashboard)
vscode-extension/ — VSCode/Cursor extension
.claude/
  agents/       — Agent definitions
  skills/       — Skill files (self-mark, self-search, etc.)
  rules/        — Shared rules

Before submitting:

1. Run pnpm test and pnpm build to verify nothing breaks
2. If adding a new hook event or MCP tool, update the relevant docs
3. Keep PRs focused — one feature or fix per PR

License

Source Available — free for personal, educational, research, and open-source use. Commercial use requires a license. See LICENSE for details.