The feedback loop is where the split gets expensive. Memory citations carry rollout IDs, and when Codex emits one, runtime code increments usage counts and timestamps on the matching Phase 1 records; later consolidation passes prefer evidence that has been used frequently and recently [7]. The ranking signal, in other words, is agent behaviour. A build command that stopped working three weeks ago keeps earning its place every time an agent cites it, while the corrected version enters with a usage count of zero.
Nothing in the described pipeline checks a remembered fact against the repository as it stands. Phase 2 receives selected evidence plus a git workspace diff [8], and the prompts ask the models to merge new evidence, rewrite stale guidance, preserve uncertainty and drop claims whose evidence has disappeared [9]. That is judgement applied to what the extraction stage happened to surface, not verification. What the working agent actually sees is narrower still: a bounded `memory_summary.md` injected into developer instructions, after which the model searches `MEMORY.md` and opens one or two supporting summaries or skills [11].
If you spot the stale fact yourself, the route back is indirect. According to the analysis, a direct request can produce an append-only ad-hoc note, but only when dedicated memory tools are enabled or the active permission profile separately permits the write, and the note changes nothing until the next consolidation pass interprets it and regenerates the artifacts [10]. That is two conditionals and a scheduler between "this command is wrong" and the file your next session reads. Dispatch is not on session start either; the author notes memory work is triggered only after a fresh turn [15].
Count the deletion paths in the source and you get seven: an extraction dropped after a later no-op, summaries falling out of the selected set, the consolidation agent removing unsupported guidance, pruning of unused database rows, expiry of extension resources, memory attached to a deleted thread, and a full reset of the store [14][5]. Not one of them fires because a fact became untrue. Six are plumbing and set membership; the seventh is a model being asked to notice. A wrong-but-cited fact satisfies none of the conditions, which is the mechanism by which last month's guidance survives into next month's sessions.
Two more things worth holding onto from this reading. First, the semantic work is a two-model background pipeline, one model extracting reusable material per eligible rollout and a second consolidating into a global file-based workspace [16], so the artifact your agent trusts was authored by a model you did not watch write it. Second, `AGENTS.md`, session persistence and compaction remain separate systems, and the analysis puts team rules squarely in explicit instructions [12]. That division is the practical control surface here: anything you cannot afford to have quietly rewritten does not belong in the pipeline that rewrites things.
