Most agent "memory" is just a vector store bolted onto a chatbot: every message gets dumped in and semantic search hopes to pull the right snippet back. memU treats memory as a compile step instead — raw chat logs, documents, code, and tool traces are restructured into typed, categorized items before retrieval, so the agent reads back curated context rather than a pile of nearest-neighbor chunks.
What Sets It Apart
- Three layers, not one bucket. It separates a navigable Index (what the agent knows), Skills (reusable patterns and tool usage learned from traces), and Memory (user profile, goals, key events) — each answers a different question instead of cramming everything into one embedding space.
- A
memorize()/retrieve()contract. Ingestion compiles sources into auditable items with source traceability; retrieval serves scoped, ranked context, so you can trace why a fact was recalled rather than trusting an opaque similarity score. - Numbers it puts forward: 92% average accuracy on the Locomo benchmark and up to ~95% token-cost reduction versus replaying full history into the prompt.
- Runs where you need it. Pluggable storage (in-memory, SQLite, Postgres + pgvector) and multi-provider LLM routing (OpenAI, DashScope, Voyage, OpenRouter), with a managed cloud at memu.so if you would rather not self-host.
Great Fit / Look Elsewhere
Great fit if you are building a long-lived coding agent or assistant that must remember project context and user preferences across many sessions, and you value auditable, structured recall over a raw RAG dump. Look elsewhere if you only need short-term context for a single conversation, want a zero-dependency one-file store, or cannot yet adopt its memorize/retrieve workflow — the structure that makes recall reliable also makes it heavier than a plain vector cache.
Where It Fits
It sits alongside Mem0 and MemGPT but leans harder into compiled, multi-layer workspaces rather than a single chronological memory stream — a better match when the agent's value comes from accumulated skills and traceable knowledge, not just recalling past turns.
