Persistent Memory Engine for AI: Auto-Extract, Update, and Forget Intelligently

Supermemory is an open-source memory and context engine that gives AI assistants persistent, evolving memory across conversations. It automatically extracts facts, builds user profiles, resolves contradictions, and forgets expired information — all through a single API or MCP server. Ranked #1 on three major AI memory benchmarks (LongMemEval, LoCoMo, ConvoMem), it is suitable for both end users who want a smarter AI companion and developers who need to add memory and RAG to their own AI products.

Current AI assistants are stateless by design: they treat each conversation as a blank slate, forgetting preferences, project context, and personal facts between sessions. Traditional RAG only retrieves raw document chunks — it doesn't understand that a user's stated preference has changed over time. Supermemory solves this by introducing a dedicated memory layer that tracks facts as first-class entities, handles temporal updates (new info supersedes old), resolves contradictions, and auto-expires temporary data. This allows the AI to remember not just what you said, but what you mean.

Supermemory bundles several integrated capabilities. The Memory Engine extracts facts from conversations, resolves contradictions (e.g., 'I love SF' then 'I just moved to NYC'), and auto-forgets expired info. User Profiles are automatically maintained per user, returning a combination of static long-term facts and dynamic recent context in one ~50ms API call. Hybrid Search combines RAG over your knowledge base documents with personalized memory context in a single query, so you don't have to choose between 'what the docs say' and 'what the user prefers.' It also includes real-time Connectors (Google Drive, Gmail, Notion, OneDrive, GitHub) and Multi-modal Extractors for PDFs, images (OCR), videos (transcription), and code (AST-aware chunking).

The project is a monorepo with a TypeScript core. It runs on Cloudflare Workers and Pages with a PostgreSQL database via Drizzle ORM, and leverages Cloudflare KV for caching. The memory engine sits at the core: when you call client.add(), it extracts facts, updates user profiles, detects temporal changes and contradictions, and stores the structured memory. On client.profile() or client.search.memories(), it retrieves the relevant profile summary and memory facts. The MCP server (apps/mcp/) provides a standardized interface for AI tools like Claude Desktop, Cursor, and VS Code. External data is ingested through connectors (Google Drive, Notion, etc.) or file uploads, then processed by the extractors into searchable chunks. Benchmarks can be run using the open-source MemoryBench tool included in the monorepo.

Supermemory can be used in two primary modes. For end users, install the MCP server into Claude Desktop, Cursor, or other MCP-compatible clients with a single command: npx -y install-mcp@latest https://mcp.supermemory.ai/mcp --client claude --oauth=yes. Once configured, the AI automatically saves and recalls memories and can inject a user profile via /context. For developers, install the npm or pip package: npm install supermemory / pip install supermemory. Then initialize the client, call client.add() to store conversational content, and client.profile() or client.search.memories() to retrieve context. The library provides drop-in wrappers for popular frameworks like Vercel AI SDK, LangChain, and Mastra.

Supermemory is most valuable for AI-powered applications where personalization and long-term context are critical: AI coding assistants that remember project conventions, customer support bots that track user history and preferences, and personal AI companions that learn from repeated interaction. It is also ideal for multi-agent systems where shared context across agents is needed. However, the project has notable dependencies on Cloudflare's ecosystem (Workers, D1/KV, R2), so self-hosting requires familiarity with that stack. While the core engine is open source under MIT, the managed cloud service at supermemory.ai adds convenience features and a hosted dashboard. Privacy-sensitive or fully offline deployments will need to self-host. The memory ontology is proprietary to Supermemory, meaning migrating to a different memory system may involve data transformation.