Supermemory

What Supermemory does

Supermemory is an AI infrastructure startup building a universal memory layer for AI agents. It provides state-of-the-art memory, RAG, user profiles, connectors, and extractors, all built-in with extremely low latency and compatibility with any model. The platform offers focused primitives for ingesting, understanding, routing, and retrieving context, powered by dynamic dreaming and a custom graph engine. It aims to give agents a durable understanding of people and entities over time, grounding answers in documents and knowledge bases, and enabling personalization and correctness as facts change. Supermemory offers a single API for memory, RAG, and extraction capabilities, with SDKs for major languages and model harnesses, and supports both managed cloud and self-hosted deployments.

Products

• Supermemory APIState-of-the-art retrieval and memory, RAG, and extraction in one API for developers and teams. Offers low latency recall and handles billions of tokens per month.
• Personal SupermemoryA single memory across all AI tools a user employs, allowing what's taught to one AI to be remembered by all. Includes a control panel, AI plugins (Claude, Cursor, Codex, OpenCode), and a Chrome Extension for one-click saving.
• Memory & Continual LearningPersistent, structured, state-of-the-art memory built as a knowledge graph using a custom user understanding model, powered by dynamic dreaming and a custom graph engine.
• SuperRAGExtraction to retrieval system that performs multimodal extraction, chunking, and retrieval without embeddings or vectors. Available as a filesystem.
• ConnectorsIntegrations to pull data from various sources like Notion, Google Drive, Gmail, GitHub, OneDrive, S3/R2, and web crawlers, keeping agent context fresh with real-time webhooks and scheduled syncs.
• MemoryBenchAn open evaluation platform for memory systems, built by Supermemory.

Key capabilities

• Dynamic dreaming
• Context cloud for agents
• State-of-the-art memory
• RAG (Retrieval Augmented Generation)
• User profiles
• Connectors
• Extractors
• Extremely low latency (<300ms recall)
• Works with any model
• Persistent, structured memory
• Knowledge graph using custom user understanding model
• Custom graph engine
• One API, every capability
• 100B+ tokens / month capacity
• Self-hostable
• SOC 2 Type II compliant
• TypeScript & Python SDKs
• Focused primitives for ingesting, understanding, routing, and retrieving context
• Real-time traversal (sub-300ms graph traversal)
• Memory Graph
• Document Retrieval
• Consumer Plugins
• Open Eval Platform (MemoryBench)
• Per-user memory graph
• Auto-built profiles and fact hierarchies
• Multimodal extract, chunk, and retrieve (SuperRAG)
• Semantic search and graph traversal
• Composable operations (re-rank, aggregate, rewrite queries)
• Real-time webhooks for connectors
• 4-hour scheduled sync fallback for connectors
• Incremental updates for connectors
• Automatic retry & error handling for connectors
• Observability
• Evals
• Multimodal by default (text, chats, PDFs, images, video, code via extractors and connectors)
• Managed cloud or self-host as a single binary
• Encryption
• GDPR compliant
• HIPAA BAA compliant

Use cases

• AI Assistants that remember users
• Persistent context and reasoning memory for chats, docs, and user data
• Self-improving knowledge bases
• Ingesting, syncing, and retrieving from any source for agents
• Building AI workspaces for SMBs
• WhatsApp marketing platforms for ecommerce brands
• AI Search & Research (Perplexity alternatives)
• Generative video editing with intelligent media libraries
• Note-taking and productivity apps with smart memory
• Personalization across sessions
• Maintaining correctness as facts change
• Grounding answers in source material
• Shipping multi-tenant products with isolated user/workspace memory

AI approach

Supermemory provides context infrastructure for AI agents, offering state-of-the-art memory, RAG, user profiles, connectors, and extractors. It builds a knowledge graph using a custom user understanding model powered by dynamic dreaming and a custom graph engine. The platform intelligently indexes raw data (text, files, chats) and builds a semantic understanding graph on top of entities, which is then traversed by agents for memory operations or retrieval. It emphasizes real-time updates, contradiction logic, and multimodal processing.

Tech named: dynamic dreaming, custom graph engine, user understanding model, knowledge graph, semantic understanding graph, SuperRAG, MemoryBench, LongMemEval, LoCoMo, ConvoMem, SWEContext

What it says sets it apart

• State-of-the-art memory and retrieval, leading major benchmarks (LongMemEval, LoCoMo, ConvoMem, SWEContext)
• Memory is a graph, not a blob store, allowing facts to update, connect, and be forgotten in real-time
• Built-in user profiles with static and dynamic context
• Memory and SuperRAG integrated into one engine, sharing the same context pool
• One API for all capabilities (API, MCP, plugins, SMFS, connectors, Company Brain)
• Multimodal by default, handling various data types via extractors and connectors
• Flexible deployment options: managed cloud or self-hosted as a single binary (including offline)
• Extremely low latency (<300ms recall), 10x faster than Zep, 25x faster than Mem0
• 2x cheaper than next-best for memory graph tokens with better quality
• Deduplicates at the token level, only billing for net-new content, making it cheaper for production agents that loop over the same context
• Comprehensive connectors with real-time webhooks and robust error handling
• Offers a Startup & Research Program with free credits and dedicated support