MAMA
Always-on companion for Claude that remembers your decisions and their evolution. Local-first memory using SQLite + transformers.js embeddings.
MAMA OS — Local AI Runtime with Connected Memory
Your AI remembers why, not just what.
MAMA OS is a local AI runtime that turns conversations, decisions, and work artifacts into a structured knowledge graph. Every choice is saved with reasoning. Every mistake becomes a reference. When AI agents connect to MAMA, they don't just answer — they understand your context and recommend with evidence.
Session 1: "Use SQLite for the project"
Session 5: "Switch to PostgreSQL — SQLite can't handle concurrent writes"
Session 12: "Should we use MongoDB for the new service?"
→ MAMA: "You switched from SQLite to PostgreSQL in Session 5
because of concurrent write issues. MongoDB has similar
trade-offs. Here's what happened last time..."
Why MAMA OS
AI agents today are brilliant but amnesiac. They forget everything between sessions. MAMA solves this:
| Without MAMA | With MAMA | |
|---|---|---|
| Memory | Forgets after session | Knowledge graph with decision evolution |
| Context | Generic answers | Answers grounded in your history |
| Mistakes | Repeat the same errors | Past failures prevent future ones |
| Knowledge | Only general knowledge | Domain-specific, accumulated over time |
Why We Pivoted
MAMA started as a multi-agent dashboard — skills marketplace, playground, agent swarm, cron scheduler. Those features still exist and work, but they're no longer the focus.
Why? Because AI agents commoditized all of it. Claude, GPT, Codex can all orchestrate sub-agents, execute code, and run tools out of the box. Building another agent framework is competing with the foundation model providers themselves.
What AI agents cannot do is remember. Every session starts from zero. That's the gap only a persistent memory layer can fill — and that's where MAMA's real value lies.
| Commoditized (still works, not the focus) | Unique to MAMA (core focus) |
|---|---|
| Multi-agent orchestration | Knowledge graph with decision evolution |
| Skills/plugins marketplace | Automatic memory extraction from conversations |
| Code sandbox | Cross-source intelligence (code + chat + email) |
| Cron scheduling | Domain-specific recall with "why" context |
The Vision: From Tool to Partner
Today: General-purpose AI agent (smart but no context)
+ MAMA: AI agent with your domain knowledge (smart + experienced)
Future: AI that understands why your team makes decisions the way it does
Tacit knowledge — the know-how that lives in conversations, failed experiments, and tribal memory — is the most valuable and hardest to capture asset in any organization. MAMA captures it automatically through a knowledge graph that preserves not just facts, but the reasoning behind every decision.
Architecture
MAMA OS (localhost:3847)
┌─────────────────────────────┐
Claude Code ────────┤ │
(Plugin hooks) │ Knowledge Graph Engine │
│ ┌─────────────────────┐ │
Telegram ───────────┤ │ decisions (nodes) │ │
Discord ───────────┤ │ edges (supersedes, │ │
Slack ───────────┤ │ builds_on, │ │
│ │ debates, │ │
Web Dashboard ──────┤ │ synthesizes) │ │
│ │ embeddings (vector) │ │
│ │ FTS5 (keyword) │ │
│ └─────────────────────┘ │
│ │
│ Memory Agent │
│ (auto-extract, evolve) │
└─────────────────────────────┘
│
SQLite (local)
~/.mama/mama-memory.db
Local-first. All data stays on your device. No cloud dependency. AI provider independent — works with Claude, GPT, Codex.
Knowledge Graph
MAMA doesn't just store facts. It tracks how knowledge evolves:
"Use JWT" (decision, confidence: 0.8)
│
├── superseded by → "Use JWT with refresh tokens"
│ reason: "Users complained about frequent logouts"
│
├── builds_on → "Add token rotation for security"
│
└── debates → "Consider session-based auth for web app"
reason: "Simpler for server-rendered pages"
Edge types: supersedes (replaced), builds_on (extended), debates (alternative view), synthesizes (unified from multiple).
This means MAMA can answer "why did we switch?" — not just "what do we use?"
Benchmark: LongMemEval
Tested on LongMemEval — 500 questions across 6 types, ~115K tokens of conversation history per question.
| System | Score | Model |
|---|---|---|
| Mastra | 94.87% | GPT-5-mini |
| SuperMemory | 81.6% | GPT-4o |
| MAMA OS | 81.5% | Sonnet 4.6 |
| Zep | 71.2% | GPT-4o |
MAMA matches SuperMemory on overall accuracy while running entirely locally with open-source components.
Packages
| Package | Version | Description |
|---|---|---|
| @jungjaehoon/mama-os | 0.14.5 | Always-on runtime with messenger gateways |
| @jungjaehoon/mama-server | 1.9.3 | MCP server for Claude Desktop/Code |
| @jungjaehoon/mama-core | 1.3.3 | Core library (memory engine, embeddings, DB) |
| mama plugin | 1.8.3 | Claude Code plugin (marketplace) |
| memorybench | 1.0.0 | Memory retrieval benchmarking framework |
Quick Start
Claude Code Plugin (simplest)
/plugin install mama
# Decisions are saved automatically via hooks
# Search manually when needed:
/mama:search "authentication strategy"
MAMA OS (full runtime)
npm install -g @jungjaehoon/mama-os
mama start # starts daemon at localhost:3847
Connects to Discord, Slack, Telegram. Web dashboard at http://localhost:3847.
Requires: Claude Code CLI or Codex CLI installed and authenticated.
MCP Server (Claude Desktop)
{
"mcpServers": {
"mama": {
"command": "npx",
"args": ["@jungjaehoon/mama-server"]
}
}
}
Technical Details
- Database: SQLite via better-sqlite3 (FTS5 full-text search + vector embeddings)
- Embeddings: Xenova/multilingual-e5-small (384-dim, quantized q8, 100+ languages)
- Search: Hybrid retrieval — FTS5 BM25 (lexical) + cosine similarity (semantic) + RRF fusion
- Extraction: Sonnet for structured fact extraction from conversations
- Transport: CLI subprocess (Claude/Codex) — officially supported, ToS compliant
Current Status & Roadmap
What Works Today
| Feature | Status | Details |
|---|---|---|
| Knowledge Graph | Production | decisions + edges (supersedes/builds_on/debates/synthesizes) + truth projection |
| Hybrid Search | Production | FTS5 BM25 + vector cosine + RRF fusion, 81.5% on LongMemEval |
| Claude Code Plugin | Production | Auto-save decisions via hooks, search via /mama:search |
| MCP Server | Production | mama_save, mama_search, mama_suggest tools |
| Messenger Gateways | Production | Telegram, Discord, Slack bots with memory integration |
| Always-On Daemon | Production | Cron scheduler, web dashboard at localhost:3847 |
| Evolution Engine | Production | Conservative supersede (overlap-based), builds_on for independent facts |
What's In Progress (v0.16)
| Feature | Status | Goal |
|---|---|---|
| Memory Agent Endpoint | Planned | HTTP endpoint for real-time hook events → auto-extraction |
| Noise Filtering | Planned | Reject greetings, internal prompts, duplicates from storage |
| Scope-Based Search | Planned | Replace topicPrefix with proper scope filtering |
| Temporal Metadata | Planned | event_date on facts for time-based search |
| FTS5 Migration | Planned | Permanent trigger SQL instead of runtime-generated |
What's Not Built Yet
| Feature | Target | Why It Matters |
|---|---|---|
| Pattern Recognition | v0.17 | Detect recurring workflows from accumulated knowledge |
| Workflow Recommendations | v0.18 | "Your team usually does X before Y" suggestions |
| Cross-Source Intelligence | v0.17 | Code decisions + team chat + email = unified graph |
| Memory Explorer UI | v0.18 | Visual graph of decision evolution |
| Enterprise Server Mode | v0.19 | Central server for team knowledge (vs personal local) |
| Domain Automation | v1.0 | Knowledge graph → automated workflow execution |
Roadmap
| Phase | Version | Focus |
|---|---|---|
| Done | v0.15 | Search quality overhaul, FTS5, evolution engine (58% → 88%) |
| Next | v0.16 | Memory agent, scope search, noise filtering — LongMemEval 95%+ target |
| v0.17 | Connector framework, cross-source memory, pattern recognition | |
| v0.18 | Control tower UI, memory explorer, workflow recommendations | |
| v0.19 | Stability, security audit, enterprise server mode | |
| v1.0 | General release — domain automation |
Development
git clone https://github.com/jungjaehoon-lifegamez/MAMA.git
cd MAMA
pnpm install
pnpm build
pnpm test # 2600+ tests across all packages
See CLAUDE.md for detailed development guidelines.
License
MIT
Last updated: 2026-04-01
Yorumlar (0)
Yorum birakmak icin giris yap.
Yorum birakSonuc bulunamadi