Supermemory
A memory service that turns chats and files into a web of facts that stays current.
by Supermemory · Built this? Edit this page on GitHub
In short
You send Supermemory chats, notes, PDFs or whole Google Drives. It keeps the originals for searching, and a “dreaming” step reads them and writes short facts, like “Alex works at Stripe”. Each new fact is linked to what it already knows: it can replace an old fact, add detail to one, or lead to a new guess. It also keeps a short profile of each person, so the AI always knows the basics.
Who it’s for
Mainly developers adding memory and document search to their apps and agents. People also use it directly through its plugins for Claude Code, Cursor, Codex and others, and its MCP server.
What it remembers
Short facts about people, projects and organisations, a profile of each (lasting facts and what's going on now), and the original documents.
- Memory lives
- Supermemory's cloud, or your own machine
- API keys
- A Supermemory key (cloud), or your own AI model key (self-hosted; none with a local model)
- Works with
- TypeScript & Python SDKs, Claude Code, Cursor, Codex, MCP clients, Vercel AI SDK
- License
- Partly open sourceMIT (repo), Apache-2.0 (SDKs)
- Price
- Free tier; Pro $19/mo
Under the hood
- Who writes memories
- Supermemory's own model reads what you send and writes the facts, in a background “dreaming” pass. You can steer it with instructions, or write facts yourself. In the plugins, the AI can also save and forget on purpose.
- Where they're kept
- Its own engine that mixes meaning search, keyword search and a graph of linked facts. Self-hosted, it all sits in one folder.
- How they're found
- By meaning across both facts and original documents, with optional reranking and linked facts. The profile comes along without a search.
- Extra AI calls
- Yes: its models read everything you add, billed as credits on the cloud. Self-hosted, those calls run on your model.
How it works
- 1You send contentChats, files or synced apps, filed under a user or project.
- 2Index itThe originals are split up and made searchable.
- 3DreamIts model writes facts and links them: replaces, extends, or infers.
- 4Keep a profileLasting facts and what's going on now, for each person or project.
- 5Hand it backThe profile, plus the facts and passages that match the question.
What stands out
- Memory as a graph that stays current: new facts replace, extend or build on old ones, and the old version is kept.
- Its guesses are marked as guesses. Facts it inferred rank lower until you approve them, and there's a review queue for that.
- One-command self-hosting that can run fully offline with a local model, on the same API as the cloud.
- A strong benchmark focus: self-reported #1 on LongMemEval, LoCoMo and ConvoMem, plus MemoryBench, an open tool to compare memory systems.
What it doesn’t do
- The memory engine's source code isn't public. You can run the server free on your machine, but not read or change how it works.
- Connectors (Drive, Gmail, Notion) and the MCP server are on the cloud only, not the self-hosted server.
- The docs don't say that using a memory makes it stronger, or that feelings add weight.
- No reminders that come back on their own.
- It's memory about users and projects, not the AI's own sense of self.
Measured against human memory
5 built (1 in depth), 3 partly, 1 not, 2 not known of the Field Guide’s 11 mechanisms
Encoding
Salience
Decides what's worth keeping, and how much, as a memory is written.
Its model decides what's worth keeping: small talk is less likely to become a lasting memory. There's no visible importance score. source
Emotion
Gives memories written with feeling more weight.
The engine is closed and the docs don't mention emotional weighting either way.
Storage
Forgetting
Lets unused memories fade over time on its own.
One-off moments fade unless they matter, and dated facts (“exam tomorrow”) expire. The docs don't say that unused memories fade. source
Interference
When a new memory closely matches an old one, notices and lets the newer one win.
When a new fact contradicts an old one (a new job, say), the new one wins in search. source
Retrieval
Retrieval
Remembering something makes that memory stronger.
Preferences said again get stronger, but the docs don't say that recalling a memory strengthens it.
Association
Links memories, so recalling one brings its neighbours.
Facts are linked as they arrive, and a search can bring back each result's linked facts too. source
Prospective
Brings a memory back at the right time without being asked.
Dated facts expire rather than come back, and the profile is always there rather than timed. No reminders are documented. source
Transformation
Consolidation
Reworks memories between sessions: strengthening, merging, reorganising.
“Dreaming” works in the background, grouping related conversations and merging, linking and organising what they say. source
Reconsolidation
Corrects a memory when it's contradicted, and keeps what it used to say.
A replaced fact is kept as history, and editing a memory saves a new version beside the old one. source
Gist
Turns many specific moments into general knowledge.
It draws new facts from patterns across many memories. They're marked as guesses and rank lower until approved. source
Schemas
Forms beliefs about people and projects, and weighs new facts against them.
Each person or project gets a profile of lasting facts and current context, but new facts aren't documented as weighed against it. source
Not every system is trying to be a human memory, so a hollow dot isn’t a flaw. It just says what the system doesn’t do. What the mechanisms are.
Supermemory’s makers haven’t shared what they’re building next, so only today shows here. Makers can add their roadmap the same way they correct an entry.
Compare Supermemory with the others, side by sideQuestions
- Who is Supermemory for?
Mainly developers adding memory and document search to their apps and agents. People also use it directly through its plugins for Claude Code, Cursor, Codex and others, and its MCP server.
- Where does Supermemory keep its memories?
Supermemory's cloud, or your own machine. Its own engine that mixes meaning search, keyword search and a graph of linked facts. Self-hosted, it all sits in one folder.
- Does Supermemory forget?
Forgetting: partly. One-off moments fade unless they matter, and dated facts (“exam tomorrow”) expire. The docs don't say that unused memories fade. source
- Does Supermemory correct a memory when it turns out to be wrong?
Reconsolidation: built, basic. A replaced fact is kept as history, and editing a memory saves a new version beside the old one. source
- Is Supermemory open source?
Partly open source. License: MIT (repo), Apache-2.0 (SDKs). Code: github.com/supermemoryai/supermemory, github.com/supermemoryai/claude-supermemory.
- Does Supermemory work with Claude Code?
Yes. Works with: TypeScript & Python SDKs, Claude Code, Cursor, Codex, MCP clients, Vercel AI SDK.
- Does Supermemory need an API key?
API keys: A Supermemory key (cloud), or your own AI model key (self-hosted; none with a local model).
- How much does Supermemory cost?
Price: Free tier; Pro $19/mo.