Updated October 2026
AI agent with long-term memory
Language models forget everything when a conversation ends. Long-term memory is what lets an agent remember your preferences, your projects and what it learned last week. Done badly, it's a black box that quietly collects things about you. Done well, it's memory you can read, correct and delete.
The types of agent memory
Agent memory is usually described in three kinds:
- Semantic: facts. "Prefers British spelling", "the Q3 report lives in Finance/Reports".
- Episodic: what happened. "Last Tuesday we drafted the pricing page and the user rejected the second version."
- Procedural: how to do things. "When filing invoices, rename them by date and vendor."
CoWork OS stores the first as typed facts (preferences, rules, decisions, commitments, corrections and more), keeps an archive of past work for the second, and records structured observations and lessons for the third.
How CoWork OS stores memory
CoWork OS keeps memory in a memory folder: plain Markdown notes in a local git repository, ~/CoWork Memory by default.
- Readable: open it in any editor, including Obsidian.
- Versioned: every change is a git commit, so you can see when something was learned and undo it.
- On your machine: nothing is stored on a CoWork OS server. You can sync it to a private git remote of your own, and add up to three read-only team memory repos.
To review it in the app, open Settings → Memory. What CoWork knows lists the notes so you can edit, pin or delete them; Sources shows where each memory came from.
Dreaming: memory that tidies itself
About once a day (or when you press Dream now), one model pass tidies the memory folder: it merges duplicates, removes stale notes, and saves things you said that weren't saved yet. Safe edits land as one commit you can undo; the rest wait in Settings → Memory → Review for you to accept or reject. Dreaming uses your configured model, up to 50,000 tokens a day by default. Dreaming
Keeping things out of memory
- Wrap anything in
<no-memory>to keep it from being saved. - Strict privacy marks a workspace's memories private.
- Private memories are never sent to a model, including during compression.
- When the agent wants to forget something on its own, it asks first.
Long conversations: context compaction
Memory is different from the context window. When a long task fills about 90% of the model's context, CoWork OS compacts the conversation into a structured summary and saves what matters to memory, so the task can continue. Context compaction
Memory frameworks vs a memory-first app
| Mem0 | Letta | CoWork OS | |
|---|---|---|---|
| What it is | A memory layer for developers building agents | A platform for building stateful agents | A desktop agent app with memory built in |
| Who it's for | Developers | Developers | People who use an agent day to day |
| Where memory lives | Your database or their hosted service | Your server or their cloud | Markdown in a local git repo |
| License | Apache 2.0 | Apache 2.0 | MIT |
Sources: Mem0, Letta. If you're building your own agent, a framework is the right tool. If you want an agent that remembers you without building one, use an app.
Memory works the same with cloud or local models. Recall is keyword-based, so no separate embedding model is needed. Memory engine documentation
Frequently asked questions
Which AI agent has long-term memory?
CoWork OS keeps long-term memory as plain Markdown in a local git repo that you can read, edit and delete. Developer frameworks such as Mem0 and Letta add memory to agents you build yourself.
Where does CoWork OS store its memory?
In a memory folder on your computer, ~/CoWork Memory by default: Markdown notes in a git repository, with every change as a commit.
Can I delete what the AI remembers about me?
Yes. Settings → Memory lists everything it knows, and you can edit, pin or delete any note, or clear a workspace’s memory entirely.
Does the memory work with local models?
Yes. Memory works the same with cloud or local models such as Ollama and MLX, and recall is keyword-based, so no embedding model is required.
CoWork OS itself is free and open source. Model providers, gateways, connected services, and account plans may charge separately and apply their own eligibility and usage rules.
Last reviewed 2026-10-23 against official documentation. Features, plans and prices change; check the linked sources before deciding. CoWork OS is not affiliated with Mem0 or Letta.