memory-review

memory-review — close the loop on the session you just had

The cheap half of Ralphy's learning loop. Recall (AGENTS.md step 0) loads memory INTO a session; this skill writes the session's lessons BACK. Between them sits invariant #18 (capture corrections the moment they land). Run this when a session winds down and something was learned but nobody asked for a full postmortem.

Workflow

1. Scan the conversation for signals, newest first:

   • User corrections — model pick, register, pacing, aspect, voice, phrasing the user changed after seeing output. The strongest signal: it cost the user a turn, and uncorrected it will cost one again.
   • Frustration markers — "stop doing X", "again?", "I already said", "why is it doing Y". First-class signals, not noise.
   • Durable preferences — anything phrased as "always / never / by default" about look, sound, structure, or workflow.
   • Discovered facts — a model filter hit, a provider quirk, a flag that behaved unexpectedly, a workaround that held.
   • Techniques that worked — a prompt pattern, a recipe, a sequence that future-you would otherwise re-derive.

2. Filter through the do-not-capture list (same as ralphy memory distill and invariant #18):

   • environment-dependent failures (missing key/binary/dep) — capture the FIX if there was one, never the failure;
   • negative tool/model claims ("X is broken") — they outlive the bug and harden into refusals;
   • transient errors a retry solved — the lesson is the retry pattern;
   • task progress, outcomes, narratives — logs/ + postmortem/ territory;
   • anything already covered by the repo (MODELS.md, guidelines, playbooks) or by an existing memory entry that does not need changing.

3. Dedupe against the store. For each survivor: ralphy memory search <keyword>. Overlap → re-note the existing slug (the store versions it up); no overlap → new slug, class-level name (no project ids, no error strings).