Memory-backed personalization can look like model improvement

Type: kb/types/note.md · Tags: agent-memory, computational-model, llm-reliability

An end user can give the same short request and receive a much better result after the system learns from earlier interaction. The product feels as though its model became smarter. But the causal improvement may instead be that memory retained and activated user-specific information missing from the request.

A better model can interpret supplied context more reliably and guess more accurately from general priors. It cannot make one of several prompt-compatible commissions authoritative for a particular user without user-specific evidence. That evidence might remain in the live conversation, be encoded in personalized weights, or arrive through an external memory system. A separately addressable memory interface is therefore an architectural choice, not the only possible carrier.

The distinction matters to builders even when users do not notice it. A stable interface allows the model and memory mechanism to be replaced separately and makes controlled attribution possible without denying their interaction. Each retained-intent input needs a source, subject, scope, and an indication that it remains applicable and authoritative; otherwise memory can confidently apply a stale or unrelated commission. The same boundary separates three failures that look identical from the outside: the intent was never retained, memory failed to activate it, or the model failed to use it.

Evaluation should fix the target commission independently, preferably through current user confirmation. Then hold the memory input fixed while changing models, and hold the model and request fixed while changing memory mechanisms. Report both the components' effects and their interaction rather than crediting every end-to-end gain to the model.

Relevant Notes: