Open Agentic Brain and Memory
Why do AI coding assistants forget everything between sessions?
Because every session starts with an empty context window. A model only knows what is in front of it right now. When a chat ends, what you settled in it is gone, unless something writes it back into the next session.
Switching tools makes it worse. Claude Code, Codex, Cursor and Antigravity CLI each keep their own history, in their own format, in their own folder. None of them can read the others. So the only thing connecting your tools is you, repeating yourself.
What does forgetting actually cost a developer?
More than the minutes spent re-explaining.
Repeated explanations
Architecture, conventions and past decisions, re-typed at the start of every session in every tool.
Decisions undone
A tool that doesn't know a decision was made will confidently reverse it: a removed library, a rejected pattern.
Fixes rediscovered
The bug that took an afternoon to understand gets debugged again, from scratch, next month.
Wasted tokens
The usual workaround is pasting long notes into the prompt. Slow, expensive, and mostly irrelevant to the task.
Lost lessons
What you learned the hard way never becomes something your tools know.
You become the glue
Every tool you add is one more place your context has to be rebuilt by hand.
Don't bigger context windows fix this?
No. A bigger window lets a model read more, not remember more. It doesn't decide what the next session should read, and it doesn't carry knowledge from one tool to another.
Filling the window is also the most expensive option. With 10,000 stored memories, pasting everything into the prompt would cost over 1.1 million tokens. The real problem isn't capacity. It's having a memory that knows what matters and delivers just that, wherever you're working.
What does a good memory for AI coding tools need?
We wrote down five requirements before writing any code.
- One memory for every tool. Claude Code, Codex, Cursor and Antigravity CLI read and write the same store.
- Already loaded when you start. The relevant context is there before you type, without you asking.
- Small and relevant. A short briefing for this project, not a dump of everything ever said.
- Measured, not guessed. If a change makes recall worse, it doesn't ship.
- Owned by you. Your decisions, code and mistakes don't belong on someone else's server.
How does Open Agentic Brain and Memory give AI tools a memory?
It runs a loop: capture, distill, rank, inject, measure.
Capture
Picks up what each tool learns, as it happens, into one memory on your machine.
Distill
Overnight, turns raw observations into a wiki it keeps revising, and drafts lessons from hard problems you solved. You approve them.
Rank
Combines search relevance with a record of what actually helped. Memories you mark helpful rise; misleading ones sink.
Inject
Loads a short briefing of what matters into every tool before you type.
Recall
Any connected tool can search the full memory when it needs something older.
Measure
A built-in evaluation quizzes the memory on your own history. Ranking changes are gated on it.
A new conversation starts with the briefing already loaded, in a fraction of a second. When the model needs more, it searches. And when you switch laptops, what your other machine learned comes with you.
Does it keep projects separate?
Yes, memory is filed by project, but nothing is walled off. Each memory is tagged with the repository it came from. Open a folder and the briefing starts with what happened there: its recent decisions, fixes and pinned notes.
Every briefing also carries a short map of your other projects, so a tool working in one repo still knows the others exist. Search can cover one project or all of them. Even a brand-new repo hears: "Other projects are still searchable."
Does a memory layer save tokens?
Yes. The model reads what was retrieved instead of everything.
Tokens read per question on LongMemEval-S, 500 questions. On LoCoMo the drop is from about 20.8k to 3.7k (same model).
That's 75% fewer tokens on LongMemEval-S and 82% fewer on LoCoMo, while keeping 97% of the whole-conversation score there. The startup briefing stays short by design, and you can see an estimate of the tokens memory has saved you.
How well does it actually remember?
We measure it on public long-term memory benchmarks, and gate every change on the results.
Updated 8 October 2026. Open Agentic Brain and Memory and whole conversation: measured on the same model (gpt-6-luna answering and judging), 1,540 questions; Open Agentic Brain and Memory reads about 18% of the tokens for 97% of the whole-conversation score. *The highest score of any memory product in the widely cited 2025 LoCoMo comparison, which used gpt-4o-mini to answer and judge; a different model, so read that one as direction, not a like-for-like race. LongMemEval-S: Claude Sonnet answering with LongMemEval's own judge prompts.
Who owns the memory?
You do. It lives on your machine and in your own private GitHub. To sync between machines, memory is encrypted with a key only you hold, then stored in a private repository in your own GitHub account.
No vendor server
We run nothing in the path. If a company disappears, your memory doesn't.
Only you hold the key
A leaked repository or a leaked token exposes nothing readable.
No telemetry
The software sends us nothing. Read the privacy page.
See all of it
Search every memory from every tool and project, in one place.
Change any of it
Edit, pin, delete and restore memories. Approve or dismiss what it learns.
Take it with you
Export everything to plain files with one command.
How is this different from hosted AI memory?
| Hosted memory service | Open Agentic Brain and Memory | |
|---|---|---|
| Where memory lives | Vendor's database | Your machine and your private GitHub |
| Who can read it | Vendor infrastructure | Whoever holds your key |
| Works across coding tools | Usually one product | Claude Code, Codex, Cursor, Antigravity CLI |
| Cost | Per-seat subscription | Free and open source |
| Setup | Sign up | One command, a key and a repo |
Hosted products are easier to start with, and that's a real advantage. The trade here is ownership: you keep the memory and the key, and you take on running it on your own machine.
Frequently asked questions
Why do AI coding assistants forget things between sessions?
Each session starts with an empty context window. Unless something writes earlier decisions back into it, the model has no record of what you settled yesterday, and each tool stores its history in a format the others can't read.
Don't bigger context windows solve this?
No. A bigger window lets a model read more in one session, but it doesn't choose what the next session should read, and it doesn't move knowledge between tools. Pasting everything in is also the most expensive option.
Can one memory work across Claude Code, Codex, Cursor and Antigravity CLI?
Yes. Open Agentic Brain and Memory captures sessions from each tool into one local store, loads a short briefing into each tool before you start, and lets any connected tool search the same memory.
Does adding memory use more tokens?
It uses fewer. The model reads only what was retrieved instead of a whole history: about 31.7k tokens per question instead of 127.6k in our LongMemEval-S runs.
Where is my memory stored?
On your machine. For sync it's encrypted with a key only you hold and stored in a private repository in your own GitHub account. There's no vendor server and no telemetry.
How do I install Open Agentic Brain and Memory?
It's open source (Apache-2.0) and available now on GitHub. One command installs it, connects your AI coding tools, creates your private memory repository and generates your key. It needs macOS or Linux, Node 22+, git and a GitHub account. Coming from claude-mem? One command brings your history across.
Say it once. Every tool remembers.
Open Agentic Brain and Memory gives Claude Code, Codex, Cursor and Antigravity CLI one shared memory that lives on your machine and in your own GitHub.
curl -fsSL https://raw.githubusercontent.com/rohithhn/Open-Agentic-Brain-and-Memory/main/install.sh | bash