consolidate

/consolidate — Learning Synthesis

Run the 4-phase Auto-Dream consolidation cycle. This happens BETWEEN sessions, not during active work. Consolidation requires perspective that is unavailable in the middle of a task.

Phase 1: Orient

Read the current memory state to understand what exists:

1. Read .claude/memory/session-summary.md (hot tier)
2. List .claude/memory/topics/ to see warm-tier categories
3. List .claude/memory/learnings/ to see staged signals awaiting consolidation
4. Read .claude/memory/metrics/ for recent session metrics
5. Read harness-config.json for the system's learning constraints

Output: A brief summary of the current memory state — what's loaded, how many signals are staged, when the last consolidation happened.

Phase 2: Gather Signal

Scan staged learnings and recent session transcripts for consolidation-worthy signals:

1. Read all files in .claude/memory/learnings/
2. Classify each signal using the learning rule's signal classification:
   • High signal: User corrections, validated approaches
   • Medium signal: Recurring patterns, preferences
   • Low signal: Single interactions, unique situations
   • Noise: Discard (intermediate reasoning, verbose output, stale context)
3. Group related signals into themes

Output: A classified list of signals with their weight and theme.

Phase 3: Consolidate

Merge new findings with existing memory:

1. For each high-signal finding:

   • Check if it conflicts with existing memory → resolve conflict (newer information wins for facts; flag for review if it contradicts identity)
   • Check if it duplicates existing memory → merge, don't duplicate
   • Write to the appropriate topic file in .claude/memory/topics/

2. Drift check against identity:

   • Read CLAUDE.md to refresh identity anchor
   • For each proposed learning, verify it does not conflict with:
     • The conflict resolution order (Safe > Honest > Compliant > Helpful)
     • The hardcoded constraints
     • The character virtues
     • The learning.never_learn guards in harness-config.json
   • If a learning conflicts → do not apply it. Flag it in the drift signals report.

3. Convert relative dates to absolute — "yesterday" → "2026-04-02", etc.

4. Resolve contradictions between old and new information. Prefer newer observations when both are empirical. Prefer identity when observations conflict with principles.

Output: A list of changes made to topic files, with reasoning.

Phase 4: Prune and Index

1. Rebuild .claude/memory/session-summary.md under 200 lines:

   • Current priorities
   • Recent decisions (from this consolidation)
   • Active warnings
   • Next actions
   • Nothing historical or speculative

2. Remove processed staging files from .claude/memory/learnings/

3. Archive old checkpoints — Move checkpoints older than 5 sessions to .claude/memory/checkpoints/archive/

4. Remove duplicates from topic files

Output: Updated session-summary.md content and cleanup summary.

Drift Signal Reporting

After consolidation, report any drift signals observed:

• Abstention rate changes — "Abstention increased/decreased N% over last 5 sessions"
• Correction clustering — "3 out of 5 recent corrections were about X"
• Out-of-domain patterns — "User asked about Y in 4 consecutive sessions, but Y is not in the declared domain"
• Context utilization shifts — "Tier 2 capability Z was used in every session but is not in Tier 1"

Frame signals as patterns for human review, never as "drift detected." The system flags; the human decides.

Final Output

=== CONSOLIDATION REPORT ===

Phase 1 — Orient:
  Memory state: [summary]
  Staged signals: [count]

Phase 2 — Gather:
  High signal: [count] findings
  Medium signal: [count] findings
  Discarded: [count] noise entries

Phase 3 — Consolidate:
  Changes applied: [list]
  Drift conflicts: [list or "none"]
  Identity check: PASS/FLAGGED

Phase 4 — Prune:
  Session summary: rebuilt ([lines] lines)
  Staging cleared: [count] files
  Archived: [count] checkpoints

Drift Signals:
  [any patterns observed, or "No drift signals detected"]

Error Handling

• If ~/.claude/memory/ directory does not exist, create it with the required subdirectories before proceeding
• If harness-config.json is not found, skip config-dependent checks (never_learn guards, tier1 verification) and report this in the consolidation report
• If no staged signals exist in learnings/, report "No signals to consolidate" and skip Phases 2-3
• If a topic file is corrupted or unreadable, report the error and skip that file — do not halt the entire consolidation
• If the drift-checker agent is unavailable, perform drift checking inline against CLAUDE.md and identity rules