RAG retrieves documents; it doesn't know that you switched jobs last month or changed your mind about a tool. The bet here is that memory isn't a static vector store but a living model of the user — facts get extracted, watched as they change, reconciled when they contradict, and allowed to decay when stale. That distinction is why it currently sits at #1 on LongMemEval (81.6%) and leads LoCoMo and ConvoMem.
What Sets It Apart
- Hybrid retrieval in one query: combines classic document RAG with a personalized memory layer, so a single call returns both relevant files and what's true about the user right now.
- Self-maintaining: it doesn't just append memories — it reconciles contradictions and auto-forgets expired facts, which keeps recall from degrading as history grows.
- Built for the request path: user-profile lookups return in ~50ms, fast enough to sit inside a single chat turn rather than a background job.
- Edge-native: runs on Cloudflare Workers/KV with PostgreSQL and Drizzle, and ships as a single binary you can run fully offline.
Who It's For
Great fit if you're building an assistant, agent, or app that needs durable per-user context and would rather call an API than hand-roll a memory pipeline — it plugs into Claude Desktop, Cursor, VS Code, the Vercel AI SDK, LangChain, and the OpenAI Agents SDK, and ingests from Gmail, Notion, Google Drive, and GitHub. Look elsewhere if you only need one-shot document Q&A, where plain RAG is simpler and cheaper, or if you can't send user data to a hosted service and don't want to operate the self-hosted stack yourself.
