A memory layer for LLMs that claims to beat OpenAI's own

What it does

MemOS sits between your LLM and its long-term recollections, offering a unified API to add, retrieve, edit, and delete memories. It structures memory as a graph rather than a black-box embedding store, supports text, images, tool traces, and personas, and lets you isolate or share knowledge across users and agents via composable “memory cubes.” The system can run fully self-hosted or as a cloud service, with plugins for Hermes Agent and OpenClaw.

The interesting bit

The project pushes a “self-evolving” model where memory matures through three tiers: L1 raw traces, L2 policies, and L3 world models, eventually crystallizing into reusable skills driven by feedback. That sounds ambitious, but the README also posts eye-catching benchmark deltas—claiming +43.70% accuracy versus OpenAI Memory and 35.24% token savings—so the team is clearly aiming to make the memory layer competitive, not just convenient.

Key highlights

• Graph-structured memory with a unified CRUD API, designed to be inspectable rather than an opaque vector dump
• Multi-modal support: text, images, tool traces, and personas retrieved together in one system
• Hybrid retrieval combining FTS5 full-text search with vector search, plus asynchronous ingestion via Redis Streams
• “Memory cubes” for isolating or dynamically composing knowledge bases across projects and agents
• MCP support for memory operations and local-first plugins that keep data on-device in SQLite

Caveats

• The self-hosted path depends on Neo4j, Qdrant, Redis, and several API keys, so it is not a single-binary deployment
• The README posts bold benchmark deltas but does not explain the evaluation methodology in the provided text
• The repository is tagged TypeScript, yet the self-hosted server path shown in the README is Python-based; the split is not explained in the excerpt

Verdict

MemOS is worth a look if you’re building agentic workflows that need persistent, cross-session memory with structured retrieval and don’t mind operating a small fleet of dependencies. If you just need a quick vector store for embeddings, this is probably overkill.