Every session starts from zero
If you run an AI agent, you already know the ritual. You open a session and re-explain the project: what it is, what was decided, what was rejected, why the build works the way it does. Some people keep this in md files — a CLAUDE.md here, a notes file there — and soon the agent is lost in an md jungle, reading stale half-truths from March alongside the things that are true today. Either way, you are the memory. You retype the same context a thousand times a day, and the agent repays you by re-proposing the approach you killed months ago, because nothing told it not to.
The models are getting smarter every month. But a model without memory meets your project fresh every morning. Without memory, they're all dumb — the intelligence is wasted on re-deriving what you already told it.
What memory should actually do
Before describing ours, here is what a fix has to pass. Three tests.
Already there when the session starts. A brief opens every session — where things stand, what was settled, what comes next — and on every prompt, the records that bear on that prompt arrive with it. No tool calls. No turns spent hunting through files. If the agent has to go looking, the memory isn't memory, it's a library with bad opening hours.
It keeps what the conversation settled. Not every half-thought on the way — what the exchange actually concluded. A memory that archives noise is a junk drawer with a search box.
It remembers decisions, not just facts. Records keep what is true now. But "why is it like this?" needs the history: every decision dated, with its reason and the option it beat. History never rewritten — a reversal is a new line, and the old way stays on record as the old way.
I tried most of the memory systems for AI agents. Each did a part of this, and none did what I needed. So I spent almost a year building the one that does.
Meet brethof-brain
brethof-brain is memory for AI agents that is already there when the session starts.
It works in layers, because a memory is more than a transcript. The full chat history — every session kept in full and searchable by keyword or by meaning. Curated records — the current truth of each project, which the memory builds itself by reading your conversations in stretches of five exchanges: it keeps what the conversation settled, reconciles new facts against old ones, and you never file anything by hand. Rules — standing instructions, yours, loaded with most messages, so every agent on every project works the same way. Playbooks — a procedure written once as markdown, run by any agent on any project the same way every time.
Under the records sits the graph of decisions: every decision dated, with its reason, what it replaced, what was tried and rejected along the way. When your agent suggests the thing you rejected back in March, the line saying so — with the day and the reason — comes back with the search. And the Ledger quietly catches the things you said you'd do, so they're reviewable instead of lost. Check it. You'll find the ideas you thought about and never implemented.
No "save my context" ritual. Nothing to remember. You start a session and say "continue," and you carry on. You can run ten projects at once, each walled off — its own records, rules, graph and Ledger — and nothing bleeds between them. One memory serves many agents: one on the front end, one on the back end, or one building and a second checking its work with the whole record in front of it. Your agent never re-proposes what you already killed, because the memory carries the kill.
Why we built it
We built it because we were the problem. We were fed up explaining the same thing to our agents a thousand times a day. We were lost in the md jungle ourselves. We started in December 2025 with md files, and almost everything from those first three months is gone — that is what md files do. In March we moved it onto a database, and every conversation since April is still there, searchable. Our own days run five to seven projects with several agents in parallel on one shared memory — this product was not tested on a demo tenant, it is proven every working day on its own makers, and it has been running since December 2025.
Two convictions drove the design. Teach your agent, don't gate it — memory should make the agent smarter, not add steps between you and it. And AI should do the work for the human. We try to do the work for you.
What stays private
One choice at signup, and it decides where your memory lives.
Local: the memory — records, archive, graph, search — lives in one container on your own machine. We process it; we never store it. When a piece of memory needs judging or a session needs its brief, your box sends that piece to our hub, and the hub keeps none of it.
Hosted: the whole memory is encrypted under a key derived from a passphrase only you set and hold. It unlocks when you start a session and locks itself after a short idle period. While locked, it is ciphertext to everyone, including us — forget the passphrase and no one can open it.
The hub processes pieces and stores none of them. What it holds durably is small and dull: your standing rules and one line per project saying what the project is. Processing runs in a secure instance in Zurich, Switzerland, in a Tier III+ datacenter. The model that does the thinking is bound by contract — the provider's commitment, not our guarantee — not to log, retain or train on what passes through.
You own the data. On hosted memory, export the whole history, delete a project, redact an exact phrase, or delete the account, from the panel; a local memory sits on your own disk, in your own hands. And because the client is source-available, not just trusting black boxes: you can read every line of it and see exactly what leaves your machine.
How to start — free
The install is the least work of any developer tool we know, because you barely do it: tell your agent "install brethof-brain." It fetches the client, wires the hooks, and checks its own work. You need no IT knowledge.
The one step a human does: sign up — email, GitHub or Google — choose local or hosted, copy the key. The install opens a window on your own screen where you paste it, and for hosted your passphrase, where your agent never sees either. That's it.
The free plan gives you 500 messages a month, no card. At the limit, the memory stops learning but keeps serving: the session brief, the per-prompt records, every search and the full history keep working — recall is never metered.
It runs on twenty proven platforms, each installed clean-machine and green in our rig: Claude Code, Codex, Qwen Code, OpenCode, Kilo Code, dsh, OpenClaw, Hermes Agent, GitHub Copilot CLI, Pi, CodeBuddy, Devin CLI, Kiro CLI, Qoder CLI, goose, Gemini CLI, Antigravity CLI, Cline, Amp and MiniMax Code.
If you were going to tell people about it anyway, we run an affiliate programme: brethof.ai/brain/affiliate/.
That's the whole pitch. Open a session, tell your agent "install brethof-brain," and the next session starts with your agent already briefed.
brethof-brain launched today, 6 October 2026. Start free at brethof.ai/brain.
Top comments (0)