Keep transcripts as evidence

A conversation may contain a project name, a deadline, a preference, an unresolved question and several ideas that were later rejected. Saving the transcript preserves the exchange. It does not tell the next session which statements remain true or which one should guide the work.

The problem appears as soon as something changes. A person gives one delivery date, corrects it later, then discusses a hypothetical alternative. A similarity search can retrieve all three passages. The model receives related text but still has to infer the current fact from conversational order and wording. That inference becomes less reliable when the correction happened in another channel.

Treat transcripts as source material. Durable memory should contain smaller records derived from sources, each with an identity, type, status and link back to the evidence. A record such as project delivery date can point to the message or document that established it, then be superseded when an authorised correction arrives.

This separation also controls what enters the prompt. A task may need the current date and the decision that set it. It probably does not need the full conversation around calendar options. Retrieval can return the compact record first and let the agent open the source if the task requires context.

Do not discard the transcript after extraction. The source is needed to verify wording, resolve disputes and improve extraction. It simply has a different job from the current-state store.