Trigger graph maintenance — decay weak knowledge, prune, merge duplicates, strengthen relationships.
consolidate(action: "run", wait_ms: 2000)The consolidate machine supports these actions:
run — trigger consolidation (optionally wait with wait_ms)
stats — aggregate graph health without triggering
query — operation-based graph inspection
traverse — BFS walk with depth/relationship filters
`wait_ms`: 0–30000 (max 30 seconds). How long to wait for completion before returning status.
Decay confidence — time-based confidence reduction. Nodes not accessed or reinforced gradually lose confidence.
Prune weak nodes — remove nodes below the confidence threshold. Eliminates knowledge that has decayed past usefulness.
Prune weak edges — remove edges below the weight threshold. Cleans up tenuous connections.
Strengthen co-activated — boost edges between nodes frequently retrieved together ("neurons that fire together wire together").
Merge similar nodes — combine near-duplicate nodes (similarity > 0.95 default). Consolidates redundant knowledge.
Promote timescale — move valuable, frequently-accessed nodes to slower decay tiers (fast → medium → slow → glacial).
Generate abstractions — create semantic summary nodes from clusters of episodic observations.
Nodes live on one of 4 decay tiers:
| Tier | TTL | Decay Rate | Purpose |
|---|---|---|---|
| fast | Hours | High | Working memory, session-local observations |
| medium | Days | Moderate | Recently reinforced knowledge |
| slow | Weeks | Low | Well-established, frequently-used facts |
| glacial | Months+ | Minimal | Core architectural knowledge, proven procedures |
Consolidation promotes nodes up tiers based on access frequency and confidence. Decay happens within each tier at its own rate.
End of every productive session
Every 4–5 heavy storage/learning iterations
After outcome learning bursts
Before critical retrievals (ensures clean, high-quality graph)
Consolidation also runs automatically during idle — manual triggers supplement this.
Read the current consolidation state and orchestrator metrics without triggering a cycle:
Resource URI: graphonomous://consolidation/logReturns consolidator state (cycle count, last run, config) and orchestrator plasticity metrics (current learning rate, churn estimate, counters). Useful for monitoring without side effects.
| Metric | Healthy Range | Problem If |
|---|---|---|
| Node count | Growing steadily | Stagnant or shrinking rapidly |
| Pruned count | > 0 occasionally | Always 0 (no decay) or very high (over-aggressive) |
| Merged count | Occasional | Never (duplicates accumulating) |
| Edge/node ratio | 0.5–3.0 | < 0.5 (isolated nodes) or > 3.0 (edge spaghetti) |
| Cycle count | Increasing | N/A (just informational) |
Consolidation can prune nodes that are linked to goals, which affects coverage scores. After heavy consolidation, consider re-reviewing active goals with review_goal to ensure coverage hasn't dropped below thresholds.
Consolidation neglect — never running consolidation, letting the graph accumulate noise
Running consolidation too frequently during active work (once per session boundary is usually enough)