Main idea

In a memory system where everything the assistant writes is fed back into the knowledge stock, a synthesis produced on Monday is reread on Wednesday as an established fact. The assistant is neither lying nor cheating: it reads what the base contains, and the base contains what it wrote itself. Nothing in the file distinguishes a sentence that came from an exhibitor met at a trade show from a sentence the assistant wrote while synthesizing three sources.

The loop closes silently. An approximation stated once becomes a premise the next time, then support for a broader conclusion — and the chain looks all the more solid for being consistent with itself.

This is why a context system maintains a strict separation between the source, stabilized knowledge, the current state, the synthesis and the deliverable: each of these registers states a different epistemic status, and merging all five amounts to losing track of what was observed.

Why it matters

It shifts the question of what makes an agent's memory good. It is usually judged on what it retains; it also has to be judged on what it can tell apart. A memory that keeps a lot but does not mark the origin of what it keeps manufactures false confirmation as it grows.

And it gives a simple test to run on any memory setup: reading an entry, can you say who produced it — a human, an external source, or the system itself?

Nuances and limits

The risk is not specific to machines: a team that cites its own synthesis document rather than the interviews that fed it does exactly the same thing. The assistant only changes the tempo, because it produces a lot and fast.

And the separation has a cost: you have to refuse to write a good formulation into the base until it has been qualified, which feels like a loss at the moment you find it.

Open questions

• How do you spot, in a base already built without this separation, the claims that rest only on the system's own earlier output?