Letta
An AI agent that keeps its own memory as files, and tidies them up while idle.
by Letta, Inc. (the MemGPT team, from UC Berkeley) · Built this? Edit this page on GitHub
In short
Letta gives an AI agent a notebook it owns. The most important pages (who it is, who you are, project rules) are always open in front of it; the rest stay on the shelf, but it can see every page title, so it knows where to look. The agent edits the pages itself and every edit is saved with its history. Between conversations, background helpers “dream”: they reread recent sessions and fold what they learned into the notebook.
Who it’s for
Developers and power users who want a long-lived agent with a memory and personality of its own. There's a coding CLI, a desktop app and a developer SDK, with plans from free up to enterprise.
What it remembers
Anything the agent decides to write down: its persona, who you are, your preferences, project rules and lessons, kept as readable files.
- Memory lives
- Letta's cloud, or your machine
- API keys
- An AI model key, or Letta's cloud
- Works with
- Its own CLI and app, TypeScript SDK, MCP tools, Claude Code (demo plugin)
- License
- Open sourceApache-2.0
- Price
- Free tier; Pro $20/mo
Under the hood
- Who writes memories
- The agent itself, as it works, plus background “dreaming” helpers between sessions.
- Where they're kept
- Markdown files in a git repo, on your machine or Letta's cloud.
- How they're found
- The most important files are always in view; the agent opens others by name, and a helper searches past conversations. Meaning-based search is an add-on.
- Extra AI calls
- Yes: dreaming and the helpers are extra AI calls, on your model key or Letta's credits.
How it works
- 1You talkThe agent reads its always-open pages every turn.
- 2It looks things upIt opens other pages or searches past chats when it needs to.
- 3It edits its notesIt rewrites its own memory files; each edit is saved in git.
- 4It dreamsBackground helpers reread recent sessions and fold in the lessons.
What stands out
- Self-editing memory: MemGPT pioneered an AI managing its own memory, the way a computer's operating system manages RAM.
- Sleep-time compute: tidying memory happens in the background, not while you wait for a reply.
- Memory as git-versioned files: readable, with full history, and portable across models and machines.
What it doesn’t do
- Nothing fades on its own and nothing gets stronger with use. What survives depends on what the AI chooses to edit.
- It's a whole agent, not a memory you add to the one you already use. Its Claude Code plugin is labelled a demo, not for production.
- Background dreaming can change memory without you reviewing it first (review is optional).
Measured against human memory
1 built, 7 partly, 3 not of the Field Guide’s 11 mechanisms
Encoding
Salience
Decides what's worth keeping, and how much, as a memory is written.
The AI decides what's worth writing down and where it goes, but there's no importance score. source
Storage
Forgetting
Lets unused memories fade over time on its own.
Nothing fades on its own. A memory goes away only if the AI edits it out. source
Interference
When a new memory closely matches an old one, notices and lets the newer one win.
A memory helper cleans out repeats, by the AI's judgement rather than old and new competing. source
Retrieval
Retrieval
Remembering something makes that memory stronger.
Using a memory doesn't make it stronger. source
Association
Links memories, so recalling one brings its neighbours.
The always-visible list of pages works as signposts to related ones. There are no automatic links. source
Prospective
Brings a memory back at the right time without being asked.
It can schedule prompts to run later, once or on repeat, and the agent can schedule them itself. That runs a task at a time; a memory doesn't come back on its own. source
Transformation
Consolidation
Reworks memories between sessions: strengthening, merging, reorganising.
Dreaming helpers reread recent conversations in the background and fold the lessons into memory. source
Reconsolidation
Corrects a memory when it's contradicted, and keeps what it used to say.
The AI rewrites memory freely, and every edit is saved in git, so old versions are always kept. source
Gist
Turns many specific moments into general knowledge.
Dreaming boils conversations down into lessons, done by the AI rather than a separate mechanism. source
Schemas
Forms beliefs about people and projects, and weighs new facts against them.
Always-open pages about its persona, you and the project shape how it reads new things, but only implicitly. 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.
Letta’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 Letta with the others, side by sideQuestions
- Who is Letta for?
Developers and power users who want a long-lived agent with a memory and personality of its own. There's a coding CLI, a desktop app and a developer SDK, with plans from free up to enterprise.
- Where does Letta keep its memories?
Letta's cloud, or your machine. Markdown files in a git repo, on your machine or Letta's cloud.
- Does Letta forget?
Forgetting: not built. Nothing fades on its own. A memory goes away only if the AI edits it out. source
- Does Letta correct a memory when it turns out to be wrong?
Reconsolidation: partly. The AI rewrites memory freely, and every edit is saved in git, so old versions are always kept. source
- Is Letta open source?
Yes, it's open source. License: Apache-2.0. Code: github.com/letta-ai/letta-code, github.com/letta-ai/letta.
- Does Letta work with Claude Code?
Yes. Works with: Its own CLI and app, TypeScript SDK, MCP tools, Claude Code (demo plugin).
- Does Letta need an API key?
API keys: An AI model key, or Letta's cloud.
- How much does Letta cost?
Price: Free tier; Pro $20/mo.