The memory layer for AI.
Preserves the continuity of your intent across sessions.
- Direct recall. A piece of information changes? The current version is returned, the history kept.
- A piece of information changes over the course of a project: an address changes, then a number, then the address again. Hundreds of messages later, the question is about the current value. Only the latest version counts.
- When a piece of information changes, AMAFOS returns the current version and keeps the history of the previous ones. No value is overwritten: each one remains available to consult.
- Multi-step deduction. Several updated facts, a single question. Your AI links them, step by step.
- Some answers fit in no single fact: they require linking several of them, each of which may have been updated. An error at the first step makes every following step wrong. Each link must therefore be served in its current version.
- Your AI chains the searches, and AMAFOS serves it the current version at each step. The reasoning stays its own, and the facts it relies on are up to date.
- Temporal tracking. The budget goes from 3,500 to 4,200 €. Every change is dated and kept.
- A project unfolds over time: a budget is revised, a deadline pushed back, a decision taken on a given date. Knowing what is true is not enough, you also need to know since when.
- Every piece of information is dated, every change kept with its date. AMAFOS returns the current value, the moment it took effect, and the versions that came before it.
- Contradiction. Two pieces of information contradict each other? AMAFOS stops and lets you decide.
- Two incompatible pieces of information rarely coexist by chance: one corrects the other, or one of the two is wrong. Only the person running the project knows which one counts. Deciding in their place means risking recording the error.
- Faced with a contradiction, AMAFOS overwrites nothing. It presents both versions and waits for your decision, and the older one remains available in the history.
The decision stays human.
An intention is not stated all at once. It becomes more precise, changes direction, sometimes goes back on itself. No project is written in a single instruction: it is built through successive adjustments, and each one weighs on the final result. A machine can detect a contradiction, but it cannot know which of the two versions reflects your intention. Human arbitration is not a brake: it is what keeps a small drift from becoming a fundamental error.
Memory serves thinking.
The more you fill a model's context, the less well it reasons. Information drowned in noise is poorly used, even when everything that matters is there. A model does not need to see everything, it needs to see the right things. Well served, a lighter model can match far more powerful ones. Performance does not come from size, it comes from what you serve.
Join the beta
MCP standard
Get an email when the beta opens.