dream

Consolidate project memory. This pass adds no new knowledge: it merges duplicates, resolves contradictions, drops what is stale and strengthens what is proven. Edit only MEMORY.md and, after the user confirms, the ## Project Learnings section of CLAUDE.md. Never touch project code.

Entry schema

MEMORY.md sections: ## Patterns, ## Pitfalls, ## Preferences, ## Architecture, ## Tools. Work with each entry in this typed form (step 6 covers how it is written back):

### kebab-case-key
- **Insight:** One clear sentence
- **Confidence:** N/10
- **Source:** command/agent/session that produced it (e.g. debug-user-tx, manual)
- **Files:** relative/path.rs, other/path.ts   (optional)
- **Date:** YYYY-MM-DD

Backfill missing fields: unknown source legacy, missing date today, missing confidence 5/10.

Pipeline

1. Collect. Read MEMORY.md; if it is missing, create it with the five section headers, report "fresh memory initialized" and skip to step 6. Read the ## Project Learnings section of CLAUDE.md (Recurring Issues, Fix Patterns, Config Conventions): it is always loaded and is the authoritative set for contradiction checks. Count entries per section and malformed entries.
2. Merge. Same key more than once: keep the latest-dated entry, fold extra detail from the older ones into its Insight, take the max confidence and delete the older ones (consolidation is the one place append-only history gets compacted). Different keys with the same meaning (compare the insights, not the strings): merge under the clearer key, noting the alias if useful.
3. Resolve contradictions ("use X" vs "avoid X", inside MEMORY.md or against Project Learnings). Higher confidence plus newer date wins; if the referenced files show which one is true today, check the code; if it is still ambiguous, ask the user and show both entries. Delete the loser and add "supersedes: " to the winner.
4. Prune. Auto-remove dead references (every listed file is gone, checked with Glob; if some survive, trim the Files list), transient leftovers (in-progress work, old branch state, one-off incidents that can't recur) and verbatim restatements of README or CLAUDE.md. For stale entries (confidence 4 or lower, older than 90 days, never re-confirmed) ask the user: remove, keep and re-date to today, or update the insight.
5. Re-rank each section by confidence, then date, both descending.
6. Write back under the cap. MEMORY.md loads at every session start, so it has a hard cap of 200 lines / 25KB and stays index-pointer style: one-line insights pointing at files, not essays. The typed fields are processing-only for steps 2-5; that metadata lives in the linked detail files, not in the index. If it is over the cap, compress insights to one line, drop obvious Files, then remove the lowest-confidence entries and tell the user what was cut. Show a before/after diff summary, not the whole file.
7. Export to CLAUDE.md (optional; confirm first). Promote an entry only when its confidence is 8/10 or higher, it survived an earlier consolidation (30+ days old or re-confirmed), it applies to the whole project rather than one file's quirk, and it fits one bullet: - **key:** insight (N/10). On confirmation, append it under the matching Project Learnings subsection and delete it from MEMORY.md; CLAUDE.md is always loaded, so keeping both wastes tokens.