What it does

Supermemory is a memory and context engine for AI agents. You feed it conversations, documents, or files; it extracts facts, tracks how they change over time, resolves contradictions, and forgets stale information. It also maintains per-user profiles (static traits plus recent activity) and runs hybrid search across both personal memories and uploaded documents. There is a consumer app, browser plugins, an MCP server for Claude/Cursor/VS Code, and a straightforward API with SDKs for TypeScript and Python.

The interesting bit

The project claims #1 rankings on three memory benchmarks (LongMemEval, LoCoMo, ConvoMem) and publishes an open-source benchmarking framework called MemoryBench so you can verify that yourself. The “user profile” abstraction is the practical hook: one API call returns both long-term facts and recent context in about 50ms, which you can drop straight into a system prompt. No vector DB setup, no chunking strategy debates.

Key highlights

• Hybrid search combines RAG over documents with personal memory retrieval in a single query
• Connectors auto-sync Google Drive, Gmail, Notion, OneDrive, GitHub via webhooks
• Multi-modal extractors handle PDFs, images (OCR), videos (transcription), and code (AST-aware chunking)
• Drop-in framework wrappers for Vercel AI SDK, LangChain, LangGraph, OpenAI Agents SDK, Mastra, Agno, n8n
• Scoped memory via “container tags” (projects) to separate work, personal, or client contexts

Caveats

• The README is heavy on feature lists and light on architecture details; the “how memory works under the hood” section is truncated in the source
• Benchmark claims are stated but not independently verified here; the project does provide tools to test them yourself
• The hosted service is free for consumers, but pricing and self-hosting complexity for the API tier are unclear from the README

Verdict

Worth a look if you are building agents that need to remember users across sessions and you would rather not operate a vector database. Skip it if you need full on-premise control or if your use case is stateless enough that a simple prompt cache suffices.