cos-consolidate-memory

consolidate-memory — the cold path

The only place destructive memory edits happen, and only under discipline. Reads the week's captures + corrections and reconciles them into the canonical Markdown. Output is a reviewable diff in batch + a changelog + a git commit — you review the diff, never write-by-write. Full algorithm: engine/methods/write-back.md §5.

Inputs

• instance/state/corrections.md (status: open)
• instance/log/runs/*.md since last consolidation (captures)
• instance/state/{open-loops,commitments,pending-questions}.md
• instance/config.md → write_back.*, autonomy.memory_tier_line

The five operations

1. Promote — recurring corrections (same tag ≥ threshold, or any strength: rule) become edits to the tag's single write-target (§3 table). This is the learning. Apply the §5 algorithm: group → threshold/rule → draft synthesized edit → diversity check → tier-classify.
2. Merge — duplicate entities collapse into one canonical file; redirect backlinks.
3. Supersede, don't overwrite — a stale fact gets valid_until stamped and the new fact appended. History is preserved; you can always see what was true and when.
4. Decay — age last_touched/confidence; flag + archive unreinforced low-confidence facts; dismiss corrections older than write_back.decay_weeks that never clustered.
5. Prune — move closed loops / resolved questions out of state/ into memory/episodic/.

Safety tiers (write-back §5.4)

• Tier 1 (auto + changelog): merges, supersedes, decay, most #process/#fact promotions.
• Tier 2 (propose to queue): any edit to core/ (identity/voice/autonomy/priorities), deleting sourced evidence, or carrying source-derived content into procedural/core. Write these to instance/queue/review/daily-brief-<date>.md as a raw diff (not a summary) for explicit approval.
• core/ budget pressure: if a promotion would exceed a core/ file's budget_chars, consolidate/supersede before adding (a char-count helper flags over-budget; it proposes consolidation, it is not a hard runtime block).

Health metric + low-volume mode (write-back §7 / §7.1)

• Compute per-tag correction rate with the deterministic count helper (group by tag + ISO week); interpret, don't hand-count.
• Below write_back.volume_floor for a tag → report "insufficient data", not a rate. Use the promotion-survival proxy (did a promoted rule yield zero same-kind corrections after?).
• Weeks 1–4 promoting only via #fact/#relationship/rule is the planned early phase.
• Mistagging check: flag any tag whose promotions repeatedly fail to reduce recurrence (a wrong cut — distinct from the #other rate).

Injection containment at the cold path (write-back §8)

• Provenance is sticky; no source-derived content reaches procedural/core without a Tier-2 approval, regardless of recurrence.
• Only restated derived facts are promoted — never raw source text verbatim.
• Bound the review surface: cap source-derived promotions per batch and segregate source-derived diffs from internal corrections in the daily brief.

Outputs

1. instance/log/maintenance/<date>.md — the changelog (every operation, with before/after + tier).
2. A git commit of instance/ (the diff is the review surface).
3. Tier-2 proposals appended to instance/queue/review/daily-brief-<date>.md.
4. The health report (per-tag rate or "insufficient data", promotion-survival, any mistagging/#other flags).
5. Backup commit per instance/.backup-instructions.md cadence.

Verification (a week of seeded captures)

• A single reviewable diff + changelog is produced.
• Nothing in core/ was edited without a queued Tier-2 proposal.
• A #fact correction superseded (not overwrote) its target; prior value still readable with valid_until.
• A strength: rule correction promoted at count 1; a sub-threshold #voice cluster did not auto-promote.
• The health report renders; a sub-floor tag shows "insufficient data".

Capture footer

End with the standard capture footer (engine/templates/capture-footer.md) — yes, even the cold path captures its own run for observability.