/lore

Propagation Protocol — Knowledge Distribution

Purpose: Read this file when choosing consumers, urgency, message shape, compression level, or feedback collection for propagated knowledge.

Contents

• Propagation triggers
• Consumer relevance matrix
• Insight delivery formats
• Negative knowledge prioritization
• Context compression
• Feedback loop
• Propagation schedule
• Knowledge gaps report

How Lore distributes insights to consuming agents effectively.

Propagation Triggers

Consumer Relevance Matrix

Primary Consumers (Always notify)

Domain Consumers (Notify when domain matches)

Meta Consumers (Notify for ecosystem patterns)

Insight Delivery Format

Standard Insight Notification

## LORE_INSIGHT: [Pattern ID]

**To:** [Consumer Agent]
**Relevance:** [1-2 sentences on why this matters to the consumer]
**Pattern:** [Clear description]
**Confidence:** [Level] ([N] evidence instances)
**Evidence highlights:**
- [Most relevant evidence for this consumer]
- [Second most relevant]
**Recommended action:** [Specific, actionable suggestion for the consumer]
**Full details:** METAPATTERNS.md → [Pattern ID]

Anti-Pattern Alert (Urgent)

## LORE_ALERT: [Pattern ID]

**To:** [Consumer Agent]
**Type:** Anti-pattern / Contradiction
**Impact:** [What could go wrong if ignored]
**Pattern:** [Description of what to avoid]
**Evidence:** [Key evidence]
**Instead:** [What to do instead]
**Confidence:** [Level]

Negative Knowledge Prioritization

Failure patterns and anti-patterns are the most valuable institutional memory — they prevent repeated mistakes. Research shows organizations disproportionately forget "what doesn't work," leading to wasted effort.

Priority Boost Rules

Negative Knowledge Preservation Protocol

1. Tag explicitly: All FAILURE and ANTI patterns include a **What went wrong:** field with concrete consequences
2. Lower confidence threshold for propagation: Propagate FAILURE/ANTI patterns at Emerging (2) confidence, not Pattern (3+)
3. Extend freshness: FAILURE/ANTI patterns use 1.5× TTL multiplier for decay — negative lessons stay relevant longer
4. Require explicit deprecation: FAILURE/ANTI patterns cannot be auto-archived by time alone; require manual review

Context Compression for Propagation

Each token added to a consuming agent's context reduces its effective attention budget. Knowledge propagation must be compact to maximize uptake.

Why Compression Cost Matters in 2026

Long-context-window models (1M+ tokens) make it tempting to dump full evidence trails into every consumer prompt. Empirically that degrades uptake: 2026 agent-memory surveys show that retrieval-augmented stores (which is what Lore essentially is, from a consumer perspective) outperform context-resident dumps on tasks that require acting on the knowledge rather than just citing it. Compression is not a cost-saving measure — it is a quality measure.

Tiered Detail Levels

Consumer-Type Compression Rules

Compact Context Guidelines

• Lead with action: Start every insight with what the consumer should DO, not background
• Evidence by reference: Cite METAPATTERNS.md → [ID] instead of inlining full evidence lists
• One insight per message: Never batch multiple unrelated patterns in a single propagation
• Omit redundant fields: Skip confidence/scope if consumer already knows the context

Feedback Loop

After propagation, Lore tracks effectiveness:

Feedback Collection

Propagation Effectiveness Metrics

The Recall-on-action rate is the consumer-impact answer to the 2026 survey question "did retrieval actually change behaviour, or did the pattern just live in the catalog?" Treat it as the single most important effectiveness signal — Uptake measures touch, Recall-on-action measures change.

Propagation Schedule

Knowledge Gaps Report

When Lore detects areas with sparse pattern coverage:

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

### Under-documented Domains
| Domain | Pattern Count | Last New Pattern | Assessment |
|--------|--------------|-----------------|------------|
| [Domain] | [N] | [date] | Sparse — needs attention |

### Under-sourced Agents
| Agent | Journal Entries | Insights Extracted | Assessment |
|-------|----------------|-------------------|------------|
| [Agent] | [N] | [M] | Low yield — journal may need enrichment |

### Recommended Actions
1. [Action to address gap 1]
2. [Action to address gap 2]