The memory layer AI vendors forgot to ship

What it does

Supermemory is a memory and context layer that sits between your AI tools and their short-term context window. It extracts facts from conversations, resolves contradictions when your preferences change, and automatically forgets expired information. The system also maintains per-user profiles—split into static facts and dynamic recent activity—and can run fully locally if you want to keep your data off someone else’s cloud.

The interesting bit

Instead of treating memory as a dumb vector search, it fuses RAG with structured memory in a single ontology. That means one API call can surface both your uploaded documents and your AI’s remembered observations about you. The README also claims a ~50ms response time for user profile retrieval, which is fast enough to inject into a system prompt without making the user wait.

Key highlights

• #1 on LongMemEval (81.6%), LoCoMo, and ConvoMem
• Hybrid search mode queries both document knowledge bases and conversational memories at once
• Auto-sync connectors for Google Drive, Gmail, Notion, GitHub, and others via real-time webhooks
• Multi-modal ingestion: PDFs, images via OCR, video transcription, and AST-aware code chunking
• Drop-in wrappers for Vercel AI SDK, LangChain, LangGraph, Mastra, and several other frameworks

Verdict

Worth evaluating if you are building agents that need persistent memory across sessions and would rather not wire up vector databases, embedding pipelines, and chunking strategies by hand. Skip it if your use case is strictly single-turn or you already have a RAG stack you are happy with.