Supermemory

Most RAG systems store chunks and return chunks. Each session begins cold, with no knowledge of who the user is, what they corrected last time, or which preferences have shifted. Supermemory replaces that append-only pattern with a knowledge graph that merges, contradicts, and forgets facts across sessions — the vendor describes this as ‘continual learning’ rather than retrieval. The core workflow is a single API that handles ingestion, retrieval, and memory updates together, with TypeScript and Python SDKs and an npx setup command the docs surface as the entry point.

The differentiating feature is the Profiles primitive: preference, behavior, and identity context that persists across sessions so the agent knows the person, not just the current conversation thread. Paired with the Filesystem primitive — a POSIX-mountable memory layer where grep becomes semantic search and any file is indexed automatically — agents built on Supermemory can access context the way a developer accesses a filesystem rather than constructing retrieval queries by hand.

Supermemory fits teams building production agents that serve returning users: personalized support bots, coding assistants that learn a developer’s patterns, or RAG applications that need user-specific context injected at retrieval time. It does not fit organizations that cannot send user memory data to a managed third-party service. The core memory engine is proprietary and cloud-hosted; the open-source components are the MCP server and plugins, released under MIT. Self-hosting the full engine requires an enterprise agreement — there is no self-hosted community path. Teams that hit that wall typically shift to a combination of a self-hosted vector database and a custom memory management layer, which means rebuilding what Supermemory abstracts.

The Connectors layer syncs Slack, Notion, Drive, Gmail, GitHub, and S3 in real time — the vendor states changes in connected tools are reflected in agent memory without manual imports. The Extractors layer handles PDFs, web pages, images, and audio, with chunking described as meaning-preserving across document boundaries rather than character-count splits. The vendor reports SOC 2 Type II certification and claims throughput exceeding 100 billion tokens per month across the platform.