/lore

Decay Detection — Knowledge Freshness Management

Purpose: Read this file when evaluating freshness, applying domain TTL multipliers, revalidating stale patterns, or deciding archive and resurrection actions.

Contents

• Freshness states
• Domain-specific freshness policy
• Decay signals
• Audit protocol
• Revalidation protocol
• Archive management
• Health score contribution

How Lore monitors knowledge freshness, detects staleness, and manages the lifecycle of patterns.

Freshness States

State Transitions

FRESH ──(30 days)──→ CURRENT ──(90 days)──→ AGING ──(180 days)──→ STALE
  ↑                      ↑                     ↑                    │
  └── new evidence ───────┴── new evidence ─────┘                    │
                                                              ┌──────┴──────┐
                                                          ARCHIVED      REVALIDATED
                                                          (removed)     (→ FRESH)

Domain-Specific Freshness Policy

Knowledge half-life varies significantly by content type. Applying uniform TTL thresholds causes premature archival of stable process patterns and late detection of stale security patterns. Use domain-specific multipliers to adjust the base freshness thresholds.

Domain TTL Multipliers

Applying Multipliers

When evaluating freshness state for a pattern:

effective_threshold = base_threshold × domain_multiplier

Example: SECURITY-ANTI-005, last evidence 50 days ago
  Base CURRENT threshold = 90 days → Effective = 90 × 0.5 = 45 days
  50 > 45 → State = AGING (not CURRENT as base thresholds would indicate)

If a pattern spans multiple domains, use the lowest multiplier (most aggressive decay) to err on the side of freshness.

Decay Signals

Primary Signals (Automatic Detection)

Secondary Signals (Requires Analysis)

Audit Protocol

Scheduled Audit (Monthly)

1. Scan all patterns for freshness state
2. Identify AGING patterns — add to review queue
3. Identify STALE patterns — flag for immediate review
4. Check for orphaned patterns — source agents no longer exist
5. Generate Decay Report

Decay Report Format

## Knowledge Decay Report — [YYYY-MM-DD]

### Summary
- Total patterns: [N]
- FRESH: [N] ([%])
- CURRENT: [N] ([%])
- AGING: [N] ([%]) — review recommended
- STALE: [N] ([%]) — action required

### STALE Patterns (Action Required)
| Pattern ID | Title | Days Since Evidence | Source Agents | Recommended |
|------------|-------|--------------------|----|---|
| [ID] | [Title] | [N] | [Agents] | Archive / Revalidate / Remove |

### AGING Patterns (Review Recommended)
| Pattern ID | Title | Days Since Evidence | Consumer Activity |
|------------|-------|--------------------|----|
| [ID] | [Title] | [N] | [Last referenced by consumer] |

### Orphaned Patterns
| Pattern ID | Missing Source Agent | Recommended |
|------------|---------------------|-------------|
| [ID] | [Agent name] | Archive with note |

Revalidation Protocol

When a STALE or AGING pattern needs revalidation:

Steps

1. Check source journals — has new relevant evidence appeared but was missed?
2. Check consumer behavior — are consumers still acting on this pattern?
3. Check domain relevance — is the technology/context still in active use?
4. Decision matrix:

Revalidation Sources

Archive Management

Archive Criteria

A pattern is archived (not deleted) when:

• STALE for > 180 days AND no consumer references in that period
• Source agent has been removed from ecosystem
• Superseded by a newer, broader pattern
• Technology context no longer exists

Archive Format

Archived patterns move to an ## Archive section at the bottom of METAPATTERNS.md:

## Archive

### [PATTERN_ID]: [Title] (Archived: [YYYY-MM-DD])
**Reason:** [Why archived]
**Original confidence:** [Level]
**Last evidence:** [YYYY-MM-DD]
**Superseded by:** [New pattern ID, if applicable]

Resurrection

Archived patterns can be restored if:

1. New evidence emerges that revalidates the pattern
2. Technology/context becomes relevant again
3. Consumer explicitly requests restoration

Restore to EMERGING confidence, regardless of original level — must re-earn confidence.

Health Score Contribution

Lore's decay management contributes to Darwin's Ecosystem Fitness Score:

Knowledge Health Score = Σ(metric × weight) × 100

Operational Monitoring Metrics

Track these alongside the Health Score to measure consumer-facing impact of decay:

These metrics are output-oriented (measuring consumer impact) rather than structural (measuring catalog state). A healthy catalog with high propagation lag still fails consumers.

Memory-Consolidation Backstop (Dreaming alignment, 2026)

OpenClaw and related 2026 agent-memory work separate ingestion from promotion into three phases — Light (raw capture), REM (reflection / clustering), Deep (commit to long-term memory). Lore's synthesis cycle is the local equivalent and must not skip the REM-equivalent reflection step:

• The Light-equivalent step is journal harvest (Builder / Scout / Mend / etc. writing into .agents/*.md). Lore reads but does not promote.
• The REM-equivalent step is knowledge-synthesis.md clustering + contradiction resolution + confidence scoring.
• Only the Deep-equivalent step — registering a PATTERN or higher into METAPATTERNS.md — writes to long-term memory.

Generative-Agents-style ablation studies show that removing the reflection phase degrades coherent multi-day planning within ~48 simulated hours. Treat a missed monthly synthesis cycle as an outage, not a backlog. When Reflection latency exceeds the threshold above, prioritise running synthesis over expanding the catalog.

Knowledge-Decay Adversarial Signals

Long-term-memory survey work in 2026 catalogs ways memory stores degrade beyond honest staleness: prompt injection that overwrites entries, poisoning that flips a high-confidence pattern's recommendation, and silent contradiction by adversarial new evidence. Lore is defensive in three concrete ways:

1. Append-only evidence list — never delete prior evidence rows; supersede via Archive only.
2. Confidence cannot rise on a single new datapoint — promotion requires diverse-context evidence per knowledge-synthesis.md.
3. Cross-agent corroboration is mandatory for ECOSYSTEM scope — no single source can promote an ecosystem-wide pattern.

These rules cost throughput but make memory store integrity auditable.