Skip to main content
Supermemory is context infrastructure for AI agents. It gives your agent memory, retrieval and user profiles through one API, and you can configure each part for your use case. These are the building blocks it ships with:
With Supermemory, your agent remembers what each user has told it and uses that to answer more personally and more consistently. Supermemory leads the LongMemEval and LoCoMo benchmarks and independent ones such as SWEContext. See the benchmark results.

​
How does it work? (at a glance)

Text, files, chats and connectors flow into Supermemory. The RAG path runs through the smart memory engine for search over documents; the user memory path builds a memory graph that tracks how facts change over time.
  • You send Supermemory raw data in any format - text, files, and chats, or connect it to the data sources
  • Supermemory intelligently indexes them using our user understanding model and builds a semantic understanding graph on top of an entity (e.g., a user, a document, a project, an organization). We call these entities a namespace (v3/v4 called it a container tag)
  • This knowledge is now traversed by the agent, and an automatic profile is built for it. The agent may now use it for memory operations or for retrieval.

​
Why add memory to your agent?

Without memory, every session starts from zero. The model cannot know what the user preferred last week, which project they are on, or that a fact has changed since yesterday. Memory gives an agent a lasting understanding of people and entities over time: their preferences, decisions, relationships and corrections. Retrieval (RAG) grounds answers in documents and knowledge bases. Most agents need both. With memory, your agent can:
  • Personalize answers with preferences, roles and history from earlier sessions, without putting the whole chat log in every prompt.
  • Stay correct when facts change. If a user says “I love Adidas” and later “I’m switching to Puma”, only the newer preference should hold.
  • Pull the right policy, ticket or document when a question needs source material.
  • Keep each customer’s memory separate, so one user’s data never leaks into another’s.
Think of memory as the context a good teammate carries in their head, not a search box over raw logs. To see where retrieval ends and memory begins, read Memory vs RAG.

​
Why Supermemory?

  • State of the art on long-horizon memory — #1 on LongMemEval, LoCoMo, and ConvoMem, plus independent benches like SWEContext
  • Memory is a graph, not a blob store — facts update, connect, and forget in real time; not nearest-neighbor chunks alone
  • User profiles built in — static + dynamic context the agent should always know, ~ready for the prompt
  • Memory + SuperRAG in one engine — personalize and ground on the same namespace / context pool
  • Every door, one store — API, MCP, plugins, SMFS, and connectors share the same memories
  • Multimodal by default — text, chats, PDFs, images, video, code via extractors and connectors
  • Run it your way — managed cloud or self-host as a single binary (including offline)
memory graph
Memory, profiles, and SuperRAG share the same context pool when you use the same isolation (namespace). Mix and match for your product! A namespace can be anything - a user, a project, team, organization, etc.

​
Next steps

Quickstart

Make your first API call in minutes

How it works

Understand the knowledge graph architecture

Comparison

vs DIY vectors, thin memory layers, pure RAG

Self-host it

One binary, zero config, fully offline

Billing & plans

Credits, SM tokens, and how usage works

Security & compliance

SOC 2, GDPR, HIPAA BAA, encryption