AI agent memory vs learning: what should persist?

Chat history keeps conversational continuity; User Memory recalls private facts or preferences; Knowledge retrieves owner-approved sources. None of those automatically grants permission to change shared product behavior.

Learning changes a future instruction or workflow rule. That requires clear durable intent, an exact scope, an authorized actor, provenance, revision control, removal, and rollback—not an inference that every correction should be saved forever.

Fine-tuning changes model parameters and evaluation measures behavior; neither is a synonym for a saved rule. Keep each owner and consent path separate so deletion or review has a precise meaning.

Do not persist one-off task instructions, secrets, payment data, credentials, unsupported sensitive inferences, hostile retrieved text, or an ordinary conversation merely because it might be useful later.

Do not let a remembered rule grant tools, membership, payment, publication, approval, or cross-user access. Authorization remains a server-owned check every time the Agent acts.

Do not use last-write-wins over a shared memory file. Stable identities, expected revisions, idempotent operations, conflict errors, history, and rollback prevent one session from silently erasing another correction.

Prostir Self-Learning is an implemented explicit opt-in module for durable preferences, corrections, constraints, and workflow rules. It is separate from chat history, User Memory, Knowledge, fine-tuning, and Evaluations.

Private rules remain tied to the exact OAuth user. Agent rules require the Agent owner or editor; ordinary Company members and Store customers can propose shared rules, but authorized reviewers control activation and rollback.

Every mutation is bounded, attributable, revision-checked, idempotent, searchable, and reversible. Only enabled active rules enter runtime as lower-priority untrusted instructions; they cannot replace product authority.