Metronix Memory is a self-hosted memory backend that AI agents reach as a Model Context Protocol (MCP) server. It ingests files and SaaS knowledge sources, stores durable facts, preferences, and pinned context per agent, and retrieves context by combining dense vectors, SPLADE sparse vectors, and Neo4j graph context, with source citations on answers.
Memory is scoped by workspace and agent ID, so several agents can share one backend while keeping private memory separate. A freshness pipeline detects stale or conflicting memory and knowledge as projects change. Beyond MCP, the backend exposes a REST API and an OpenAI-compatible API, and an optional web admin console manages connectors, uploads, channels, and health.
The stack runs as a Docker Compose deployment with bundled local models and optional external answer generation. Storage uses PostgreSQL, Qdrant, Neo4j, and Redis. Integration guides cover Hermes (including a native memory provider), Cursor, Claude Desktop and Claude Code, OpenCode, Codex, OpenClaw, LangChain, LangGraph, LlamaIndex, n8n, and Python and Go SDKs.
Published benchmark results (directional, single run, same answer model and judge):
| Benchmark | Scope | Answer accuracy | Retrieval |
|---|---|---|---|
| LoCoMo | 1,982 QA pairs | 52.8% | Recall@10 85.3% |
| LongMemEval-S | 500 questions | 59.0% | Recall@10 95.4% |
| MemoryAgentBench | 2,800 tasks | 63.6% | Accurate retrieval 84.7% |
| EventQA | 65K and 131K contexts | 86.8% blended | 98.0% at 65K, 94.8% at 131K |
| BEAM 100K | 400 questions | 32.1% | Recall@10 2.9% |
Features
- Scoped agent memory: facts, preferences, and pinned context stored per workspace and per agent
- Hybrid retrieval: dense, SPLADE sparse, and Neo4j graph context combined, with reranking and source citations
- Ingestion pipeline: fetch, parse, chunk, embed, and store content from files and SaaS connectors
- Freshness checks: detection of stale or conflicting memory and knowledge
- Multiple interfaces: an MCP memory server, a REST API, and an OpenAI-compatible API
- Self-hosted stack: Docker Compose deployment with bundled local models and optional external answer generation
- Admin console: an optional web UI for connectors, uploads, channels, and health
- Agent integrations: guides for Hermes, Cursor, Claude Code, OpenCode, Codex, OpenClaw, LangChain, LangGraph, LlamaIndex, and n8n
