Capture what happened
Conversations and connected sources provide evidence. Selected material can become beliefs, goals, preferences and procedures. The original record and the interpretation serve different purposes.
For example, “the supplier delivered late twice” is a claim to support with records. “Avoid that supplier for urgent work” is a possible lesson. The lesson should retain its reason and remain open to correction.
Keep a stance, not just a sentence
A Mind can hold a claim with an owner, source, confidence and importance. Another Mind can hold a different view of the same situation.
Confidence is an estimate, not a guarantee of truth. Importance is about what deserves attention. A highly important claim may still be uncertain; those two signals should not be collapsed.
What isn’t used can fade
A belief nobody brings up loses prominence over time, unlike a filing cabinet where everything sits at equal weight until someone deletes it. Explicitly confirming an important belief keeps it from fading below a floor you set; locking one keeps it from moving at all.
Fading is not deleting. A faded belief can still be found and reconsidered, just not surfaced ahead of what is active or recently confirmed.
Select context for the task
The current message, recent conversation, participants and work in progress help establish the situation. The system supplies a bounded set of relevant Mind material rather than pouring the entire history into every prompt.
Exact history remains a separate source of evidence. Retrieve it when the task needs an old quote, identifier or detail. Selection and retrieval are engineering problems with measurable misses, not perfect recall.
Reason, act and verify
The model uses the supplied context and available tools to work through the task. It may need fresh information before making a recommendation. Knowing last year’s price does not establish today’s stock, costs or delivery date.
A suggestion, an authorized action and a verified result are three different states. Keep that distinction visible in the workflow.
Use corrections and outcomes
A correction can update the context used later. Repeated experience can inform future judgments without rewriting the model’s underlying weights. In this site, learning means changing the durable context and processes around the model.
The test is the next relevant task: did the system retrieve the lesson, apply it appropriately and avoid the same mistake? Completing a memory update alone does not prove improvement.