fix(consolidation): reduce decay rate, add importance floor, filter archived memories (#78)

[jack-arturo]

Summary

Fixes the overly aggressive memory decay formula that destroyed 84% of production memories within 30 days, regardless of importance.

The problem: base_decay_rate=0.1 (exponential) meant a 30-day memory had exp(-3) = 0.05 relevance, and a 90-day memory had exp(-9) = 0.0001. Production audit showed:

• 84.3% of 9,318 memories classified as "archive" (relevance < 0.05)
• 2,736 high-importance (≥0.7) memories with relevance < 0.1 ("unfairly killed")
• 409 protected types (Decision/Insight) with near-zero relevance
• Average relevance score: 0.049

The fix:

• Decay rate: 0.1 → 0.01 (configurable via CONSOLIDATION_BASE_DECAY_RATE), changing the half-life from ~7 to ~69 days
• Importance floor: relevance = max(calculated, importance * 0.3) — a memory with importance 0.8 can never decay below 0.24. Configurable via CONSOLIDATION_IMPORTANCE_FLOOR_FACTOR
• Archive filter: Added coalesce(m.archived, false) = false to keyword search (was only filtered in trending). Added post-filter for Qdrant vector search. Consolidated archived flag syncs to Qdrant payload.
• Rehabilitation scripts: scripts/audit_relevance.py (read-only damage assessment) and scripts/rescore_relevance.py (one-time recalculation with --dry-run and --target railway options)

Deployment safety: Delete/archive thresholds remain at 0.0 (disabled). After merge and deploy:

1. Run rescore_relevance.py --target railway --dry-run to preview
2. Run rescore_relevance.py --target railway to rehabilitate (avg score: 0.049 → 0.376)
3. Monitor for 1 week, then enable conservative thresholds

Benchmark Results

Benchmark	Baseline	This PR	Delta
locomo-mini (304 Qs)	76.97%	76.97%	+0.0%
locomo-full (1986 Qs)	80.06%	79.51%	-0.55% (within variance)

No regression — expected since benchmark snapshots use fresh data unaffected by production decay.

Test plan

• Unit tests pass (make test — 110 passed, 11 skipped)
• Consolidation engine tests updated for new formula
• locomo-mini: 76.97% (no change)
• locomo-full: 79.51% (-0.55%, within variance)
• Rescore dry-run against production: avg 0.049 → 0.376, 9,257 improved
• Production backup taken before any changes
• Post-merge: run rescore on production
• Post-rescore: monitor decay health for 1 week
• After 1 week: enable thresholds (CONSOLIDATION_DELETE_THRESHOLD=0.01, CONSOLIDATION_ARCHIVE_THRESHOLD=0.05)

Closes #78

Made with Cursor

[coderabbitai]

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

📥 Commits

Reviewing files that changed from the base of the PR and between 14bed1e and 97e50f5.

📒 Files selected for processing (12)

• CLAUDE.md
• app.py
• automem/config.py
• automem/consolidation/runtime_bindings.py
• automem/consolidation/runtime_helpers.py
• automem/search/runtime_recall_helpers.py
• benchmarks/EXPERIMENT_LOG.md
• consolidation.py
• docs/ENVIRONMENT_VARIABLES.md
• scripts/audit_relevance.py
• scripts/rescore_relevance.py
• tests/test_consolidation_engine.py

_{✏️ Tip: You can disable in-progress messages and the fortune message in your review settings.}

✨ Finishing Touches

• 📝 Generate docstrings (stacked PR)
• 📝 Generate docstrings (commit on current branch)

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment
• Commit unit tests in branch exp/78-decay-fix

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}