# Memoria

> The World's First Git for AI Agent Memory. Snapshot, branch, merge, and rollback your AI agent's memory — persistent, version-controlled, and MCP-native.

Memoria is open-source infrastructure (Apache-2.0) that gives AI agents a persistent, version-controlled memory layer. It is backed by MatrixOne's native Copy-on-Write engine and accelerated by NVIDIA cuVS for GPU-powered vector search.

## What Memoria Does

- **Persistent Memory**: Memories survive across sessions. Agents never start from zero.
- **Git-Level Version Control**: Every memory mutation is snapshotted. Branch, diff, merge, and rollback — for memory, not code.
- **Cross-Agent Sharing**: One agent stores a memory; every MCP-compatible agent instantly shares it.
- **Semantic Retrieval**: Hybrid vector + full-text search retrieves by meaning, not keywords.
- **Self-Governance**: Automated contradiction detection, low-confidence quarantine, and deduplication.
- **Full Auditability**: Complete provenance chain on every store, retrieve, and rollback.

## Key Metrics

- ~70% token savings via precise recall instead of full-context injection
- 12× GPU-accelerated index builds via NVIDIA cuVS (CAGRA algorithm)
- 10× query throughput improvement over CPU baselines
- 95% recall rate with sub-millisecond latency at scale
- 5-minute setup via MCP configuration

## Supported Agents

Kiro, Cursor, Claude Code, OpenAI Codex, Gemini CLI, OpenClaw, and any MCP-compatible agent.

## Memory Types

Memoria supports six memory types with different granularity and lifecycle:
- `semantic` — facts and knowledge
- `profile` — user/agent preferences
- `procedural` — workflows and instructions
- `working` — short-lived session context
- `tool_result` — cached tool outputs
- `episodic` — past interaction records

## URLs

- Homepage: https://thememoria.ai/
- App / Dashboard: https://thememoria.ai/app
- GitHub (source): https://github.com/matrixorigin/Memoria
- Quick Start: https://github.com/matrixorigin/Memoria#quick-start

## Contact

- Email: memoria@matrixorigin.io