The problem isn't that your AI forgets. It's that it doesn't go looking.
Stéphane Commenge, with Claude ·
The Fact
Nearly all the AI assistants on the market have a memory: they retain preferences, facts, summaries of past conversations.
Yet on a project lasting several months, they sometimes answer without the decision made three weeks earlier, or suggest again an option that was already ruled out. The information exists somewhere. By default, the AI doesn't go looking for it.
The Signal
For a long time we thought the problem with AI was forgetting. Give it a memory, and everything will get better.
In practice, a memory that accumulates isn't enough. A replaced decision sits next to the one that replaced it, an abandoned option stays there, and it's up to you to sort it out.
What matters is not the amount of memory. It's that the right information arrives at the right moment, and that what is no longer true stops being served.
The Action
For your longest project, this week:
1) Open a 'project file' document.
2) Write in it under four headings: decisions in force, pitfalls already encountered, options ruled out and why, next actions.
3) Give it to your AI at the start of each session, and update it at the end.
You'll see the difference from the second session.
Your AI has a memory, but it doesn't go looking
If you work with an AI on a project that lasts, you've surely lived this scene. You settled a question three weeks ago. You told the AI clearly. Today, it suggests exactly the option you had ruled out, with great confidence.
It hasn't "forgotten" in the strict sense. The decision may be in its memory, in an old conversation, in a project document. But at the moment of answering, it doesn't have it in front of it, and nothing pushes it to go and find it. It answers with what it has, that is, with what you have just told it.
That's the real problem with AIs on long projects: not forgetting, the missing reflex.
Why AI memory isn't enough
The memories built into assistants have made real progress. They retain your job, your style preferences, the broad outlines of your projects. For everyday conversation, that's very useful.
On a file that lives for several months, two limits appear.
They accumulate, they don't sort
A memory that retains everything ends up containing contradictory things. The June decision and the September one that replaced it. The price considered at the start and the price finally chosen. Without a clear mark of what has stopped being true, the AI may pick either one.
They don't distinguish the nature of information
A preference, a decision, a mistake not to repeat, an idea tested then abandoned: these are not information of the same nature. A good project manager doesn't file them in the same place. A free-text memory mixes them.
The result is familiar to everyone who runs projects with AI: you spend a huge amount of time re-explaining, correcting, reminding. The promised productivity gain erodes in repetition.
What a good colleague does, and what the AI should do
Think of an experienced colleague taking over a file. Before proposing anything, they open the file. They look at what was decided, what was tried, what failed and why. They don't propose the abandoned option again, because they know it was abandoned.
This behavior rests on two things: a well-kept file, and the reflex to open it at the right moment. Most current uses of AI have neither.
The good news is that you can keep the file starting today, without any particular tool.
Keeping a project file the AI consults
Four headings are enough
Decisions in force: what is settled, with the date and the reason. Pitfalls: what has already cost time and mustn't be rediscovered. Options ruled out: what was tried or proposed then abandoned, and why. Next actions: what remains to be done, the most urgent first.
This structure changes everything. A replaced decision doesn't stay next to the new one: you remove it, or move it to the options ruled out. The AI receives an up-to-date state, not an archive.
Open and close each session
At the start of a session, give the file to the AI. At the end of a session, ask it to propose the update of the four headings, and review it. Five minutes per session, which save thirty on the next one.
To check the quality of this update, a second conversation with no context can review it, using the produce, attack, correct method.
Write sentences that stand on their own
"He said yes" is useless in three weeks. "The client approved the €12,000 quote on September 12" is useful to everyone, including another AI that has never seen the conversation.
Why I'm building Mycelora
I've kept this project file for months on all my projects. Doing it by hand works, but it takes a discipline nobody sustains over time. That's why I'm building Mycelora, a project memory for AI assistants, hosted in France.
The principle is the one described above: memories sorted by type (decisions, refuted options, landmarks), a project sheet updated at the end of the session, and a handover to pick up where you left off. The goal is for the AI to have in front of it what matters for the question of the moment, without you having to remind it.
I use it every day on my own projects, and it's that use that is what produced the tool, as with the other apps I've built. The subject of this article isn't the tool, though. It's the method: a file that is kept, consulted at the right moment, cleared of what is no longer true. It works with a simple document, and it works better when a tool takes the burden off you.
In practice, how to get started
- Choose your longest project, the one where you re-explain the most often.
- Create the file with four headings: decisions in force, pitfalls, options ruled out, next actions.
- Fill it in a first time by asking the AI to reread your latest exchanges and propose the content. Correct it.
- Give it to the AI at the opening of each session, and have it updated at the close.
- Remove what is no longer true, instead of piling it up. It's the simplest rule and the most forgotten.
Your AI doesn't need to remember everything. It needs to have in front of it what matters, at the moment it matters.