You are your AI agent's memory.
That ends today.

The first thing you type to your agent most mornings is an apology wrapped around a briefing. "Right, remember, we're using Postgres now, not SQLite — I told you that last week — and the auth refactor isn't merged yet, and when I say 'the migration', I mean the one from March, not the one I killed." You are not prompting. You are being the memory. The AI got smarter; the part that remembers stayed a clipboard you carry by hand.

If this is your day, today is the day it changes. This article is about what the ritual costs you, why the fixes you've already tried don't fix it, and what's different about a memory that arrives with every prompt without being asked — which is what we're launching: brethof-brain.

A day as your agent's memory keeper

It starts at session open. The agent has no idea what the project is, so you spend the first ten minutes re-explaining a thing you explained yesterday. Then it asks how deployments work, and you type the same three sentences you typed on Tuesday. By mid-morning you're pasting context blocks by hand — the schema, the constraints, the naming conventions — into prompts, because you learned the hard way that if you don't paste it, it isn't there.

Then the ritual reverses. The agent proposes Redis for a queue, and you type "as I mentioned earlier, we tried that in March and it didn't hold up" — and you keep typing it, because mentioning earlier is not the same as it being known. By afternoon you're the search engine for your own decisions: which option we picked, why we picked it, what we rejected. By evening, if you're honest, you've explained the same project to your own agent somewhere north of a thousand words, most of it a repeat.

One of these days is an annoyance. Two hundred of them is a tax on every project you run. We were fed up explaining the same thing a thousand times a day, so we built the thing that stopped it. Before we get there, it's worth being precise about why nothing you've already tried fixes it — because the failures have a shape.

Why the usual fixes don't fix it

The md-file jungle is the first attempt everyone makes. A NOTES.md, a DECISIONS.md, a CLAUDE.md you grow until it's a junk drawer. Two problems: you're still filing by hand, and the file only helps if the agent reads the right part at the right time — which it doesn't, because nothing tells it which part bears on the prompt in front of it.

Session-start loads are the refinement. A memory file loads once when a session opens, and by prompt three it's buried under everything the session has said since. Other memory tools wait to be asked — the agent must decide to call them. Waiting to be asked is the whole problem. You're already the one being asked; that's the job you're trying to quit.

The diagnosis, then: the missing thing is not a better file or a bigger load. It's memory that arrives on its own — before you ask, every prompt — with the records that bear on what you just said.

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.

The difference: memory that arrives before you ask

Here is what the day looks like when the memory does the remembering. You start the PC, open the session, and the agent is already briefed: the standing rules, what each project is, the notes from the last session, the next lines of the plan, an index of what's known. You type "continue" and it continues. That sentence used to be a fantasy; it's the default now.

On every prompt after that, the records that bear on it arrive with it. The agent spends no tool calls, no turns, no hunting — the memory is simply there when it answers. It doesn't dump everything in: a session opens with an index of your playbooks, not their text, and a prompt gets what bears on it and nothing else. Ask for more when you want more.

And when the agent reaches for something you killed in March, the line saying so — with the day and the reason — comes back with the search. The graph of decisions keeps every decision dated, the option it beat, and what was tried and rejected along the way, all in the words of the conversation that settled it. History is never rewritten; a reversal is a new line. "Why is it like this?" finally has an answer.

The short form: it answers before you ask.

What the Brain is

brethof-brain is the memory that does this. It keeps, together:

• Curated records — the current truth of each project, kept by the memory itself, reconciling new facts against old ones.
• The full chat history — every session in full, searchable by keyword or by meaning.
• Your rules — loaded with most messages, so your agents work your way everywhere.
• The decision graph — how things got to be the way they are, dated and reasoned.
• Notes — the handover between sessions.
• Playbooks — a procedure written once as a markdown document, run by any agent on any project.
• The Ledger — it catches the things you said you'd do, so you can review what you missed and turn the keepers into goals your agents run to completion.

You don't file anything. You don't run a "save my context" ritual. The memory curates itself — it reads conversations in stretches, so it keeps what the conversation settled, not every half-thought on the way. Ten projects run side by side with nothing bleeding between them: each has its own records, its own rules, its own graph, and its own Ledger, and on hosted memory a key scoped to one project cannot see or search another.

Getting it, without the wall of commands

Every developer tool you've installed this year handed you a wall of commands to type. This one you ask for. Tell your agent:

install brethof-brain

That's the install. The agent fetches the client, wires the hooks, and checks its own work. The human steps are the ones every application on earth asks for: sign up, choose local or hosted, set your passphrase, and paste the key into the connect window that opens on your own screen — the agent never sees it.

Local: one container on your own machine; your memory is stored there and nowhere else — we process it, we never store it. Hosted: no server to run, no config files, and the memory is encrypted under a key only you hold — forget the passphrase and no one can open it, including us.

It works with the harnesses we've proven it on, one real install at a time on clean machines: Claude Code, Codex, Gemini CLI, and seventeen more across Linux and Windows — the full proven list is on the site. Same memory, same behaviour, whichever you use.

Try it and stop being the memory

Start free, connect one project, and run one real day. Count how many times you re-explain something. That number is the whole argument.

The founder's version, kept as his words: teach your agent, don't gate it. All the AI is getting smarter — but without memory, they're all dumb.

And one line for the people who were going to talk about this anyway: we run an affiliate programme — recurring, first-party tracked, through brethof.ai/brain/affiliate/.