A hippocampus for your agent swarm

What it does

MindMemOS sits between your agents and their forgotten conversations. It stores user profiles, preferences, and task history as searchable memories, then injects relevant context into agents before they act and writes new interactions back afterward. Over time, it consolidates what happened offline and promotes frequent patterns into reusable skills. It exposes a FastAPI service and Python SDK, and you can run it locally or use the hosted cloud version.

The interesting bit

Instead of locking memory inside one agent framework, MindMemOS treats long-term memory as infrastructure: OpenClaw, DeepSeek Harness, Claude Code, and OpenHands can all read from and write to the same store. The unusual twist is a bidirectional pipeline where raw experience memories distill into skill candidates, while skill execution traces and feedback flow back into memory to drive further evolution.

Key highlights

• Portable across agents via plugins for OpenClaw and DeepSeek Harness, with a common HTTP API and Python SDK
• Self-evolving system that learns memory schemas, runs offline “dreaming” consolidation, and optimizes retrieval through feedback
• Experience-to-skills pipeline: memories graduate into skill candidates, and skill outcomes feed back into the memory graph
• Local deployment stack includes Qdrant, Neo4j, Kafka, and optional ClickHouse/Grafana observability
• Cloud-hosted option available if you do not want to manage the backend yourself

Caveats

• Local setup is not lightweight: you must configure separate chat, embedding, and rerank model routers and spin up multiple Docker services
• The plugin ecosystem is currently limited to OpenClaw and DeepSeek Harness; broader framework support is listed as in progress
• The README mentions “dreaming” and schema learning but does not detail the mechanics, so the black-box factor is unclear

Verdict

Worth a look if you are running multiple agent frameworks and need persistent, cross-agent memory with a path toward automated skill discovery. Skip it if you want a quick, single-agent memory drop-in or lack the infrastructure appetite for Qdrant, Neo4j, and Kafka.