Memory Engineering Is a Promotion Pipeline, Not Notes
An AI agent framework runs memory through a promotion pipeline with evidence and approval instead of direct writes; rejection counts as a healthy outcome.
The memory system in ai-assistant-dot-files deliberately slows down what learned agents can write into the knowledge base, following a lifecycle of Capture -> Candidate -> Audit -> Approve -> Index -> Retrieve -> Expire. A finding first becomes a Candidate Record with required fields like source, evidence, tags, and an expiration condition; only after a memory-engineer audit and human approval does it become a durable Knowledge Item, ADR, or rule change. The design intentionally mirrors code review, since durable memory changes future agent behavior.
Perhaps the most notable trait is that rejection is treated as a healthy outcome rather than a failure: one-off, already-covered, or overly speculative findings are deliberately not promoted. Expired Knowledge Items aren't deleted either — they move to an archive with a note, because even a wrong memory is evidence of what the team once believed. A full LightRAG retrieval integration is intentionally left unimplemented for now, since the current four portable Knowledge Items don't justify the added complexity; the runbook exists so a real need has a documented, reversible path forward.
This fits into a broader governance model where all 24 agents have a clearly stated check — a structural contract, downstream review, human approval gate, or aggregate metric. The takeaway for engineers: block agents from writing directly to durable memory, require evidence and expiration conditions, treat rejection as a valid outcome, and periodically compress duplicates — because the hard part of memory isn't remembering, it's staying worth remembering.