Supermemory

Memory and continual learning for agents

Overview

About Supermemory

Supermemory is a memory engine for AI agents, offered as a hosted API, as plugins for Claude Code and Codex, and as an MCP server. You send it the raw material of a user or a task: files, chat history, URLs, documents. Its own model, called learner-1, extracts what matters, merges it with what it already knows, infers connections and forgets what is superseded, and stores the result in a vector-graph database with a sense of time. When your agent runs, the relevant memories are pulled back and injected into the context window. The problem it solves is the one every agent hits after the first session: models do not learn between conversations, and stuffing transcripts into the prompt is expensive and gets worse as history grows. The company's framing is that memory should live outside the model provider so it carries across models and harnesses, which is a reasonable position for anyone who does not want their agent's accumulated knowledge locked to one vendor.

The pricing is published and starts free. Every plan comes with a monthly credit balance drawn down at the same usage rates, and only unique content is billed; re-uploading the same document or re-syncing a connector costs nothing the second time. There is an open-source repository under MIT with more than 30,000 stars, but read the fine print before assuming you can run everything locally: the pricing page says self-hosted deployments are available on the Scale and Enterprise plans, so what the repository gives you and what the hosted learner-1 model gives you are not the same thing. The homepage also carries a set of benchmark claims, first place on LongMemEval, LoCoMo and ConvoMem and a strong SWE-ContextBench result, that are self-reported; treat them as the vendor's case, not an independent finding.