Skip to main content

Memory Routing (v3.3+)

Pluggable judgment enrichment for orchestrator agent selection, thinDispatch scoring, researcher votes, and per-task feedback. Not session continuity — that is Station vs Repertoire 0.1.

Repertoire is preferred. Repertoire (@0xray/repertoire@0.2.9 vendored) ships on. Dest is named invariants (factory + stack laws), not hangar repo-* or git slugs. Station is the compact ticket, not a substitute. Explicit opt-out only: "enabled": false, "provider": "repertoire". Scoring rules (heat is not conviction, what counts as a hit) are memory-scoring. The orchestrator links that file and does not copy the rules.

Overview​

features.json (memory_routing)
 │
 ▼
provider-loader.ts ──► createMemoryRoutingProvider()
 │
 ├── ExecutionPlanner (enrichTasks, getTaskConfidence, selectAgent)
 ├── thinDispatch.scoreAndRoute (resolveThinDispatch)
 ├── Researcher analyzeProposal (researcher-confidence.ts)
 └── ingestFeedback (per-task, v3.3)

Configuration​

In .xray/features.json or xray/features.json (validated by features.schema.json):

{
 "memory_routing": {
 "enabled": true,
 "provider": "repertoire",
 "module_path": "node_modules/@0xray/repertoire/dist/provider/memory-routing-provider.js",
 "config": {
 "signalsPath": ".xray/state/repertoire/curated_signals.json",
 "statePath": ".xray/state/repertoire/inference-state.json",
 "feedbackDir": ".xray/state/repertoire/feedback"
 }
 }
}

Provider options​

providerUse case
"null"Disable enrichment (consumer default without Repertoire)
"repertoire"Repertoire MemoryRoutingProvider
"custom"Your own createMemoryRoutingProvider() export

Disable​

"memory_routing": { "enabled": false, "provider": "repertoire" }

Full field reference: features.json.

Pipeline integration​

SurfaceMethodBehavior (v3.3–3.3.1)
ExecutionPlannerenrichTasks, getTaskConfidence, selectAgentComplexity boost, trap hints, signal-aware agent pick
thinDispatchresolveThinDispatchScore adjustment; architect override on high-confidence traps
ResearchergetTaskConfidence via provider syncMEMORY_ROUTING: block in governance output
AsideContextbuildInheritedContext → inheritedContext.memoryRouting on spawnAside
FeedbackingestFeedbackPer-task outcome loop (not aggregate-only)

Confidence gate (v3.3.1)​

ExecutionPlanner.calculateTaskComplexity() calls getTaskConfidence() when the provider implements it:

  • Adds complexityBoost from matched primitives
  • Detects highConfidenceTrapPresent / ontologicalTrapDetected
  • Passes trap context to selectAgent() for trap-capable routing

Example: trap tasks score higher complexity than plain tasks at the same estimatedComplexity.

Lessons, caps, and one plate​

A grade that passes appends one lesson on the named law. The same task id does not step the average again. LESSON_LINE_CAP is 20: past that window the oldest line leaves only after its task id is in the ledger. RETAINED_LESSON_ID_CAP is 20, insertion order, oldest first. Speech that names no stored signal mints one proposed signal tagged learned. The same speech does not mint twice. LEARNED_SIGNAL_CAP is 24 and evicts only signals that are both proposed and learned. Factory signals stay.

A plate is the pipeline schematic. Lessons are episodes on a signal. Station keeps the short subsystem table and one pointer line, Plate: <id> — .xray/state/plates/<id>.md, when intent names one pipeline. Saying goggles, suit plate, kits, host pack, or glossary recalls that plate. Those plates are domain models, not pipelines. Pre-compact and the MEMORY_ROUTING inject carry that one schematic, not the whole plate set. Stamps live in Plates. The npm package ships docs-site/docs/plates/, so a worn suit resolves the stamp from the installed reader.

Repertoire MCP (external hosts)​

In-process routing uses MemoryRoutingProvider. External LLM sessions (Hermes, Grok) should use repertoire-mcp:

"repertoire": {
 "command": "npx",
 "args": ["-y", "@0xray/repertoire", "mcp"]
}

Tools: repertoire__get_task_confidence, repertoire__get_high_confidence_signals, repertoire__search_primitives, repertoire__ingest_feedback.

See Repertoire Integration for full setup.

Custom providers​

Export from your package:

export function createMemoryRoutingProvider(
 config?: Record<string, unknown>
): MemoryRoutingProvider

Implement the contract in src/memory-routing/types.ts. Repertoire is the reference implementation.

Testing​

npm test -- src/__tests__/unit/memory-routing-integration.test.ts
npm test -- src/__tests__/unit/memory-routing-provider.test.ts
npm test -- src/__tests__/unit/researcher-confidence.test.ts
npm test -- src/__tests__/unit/researcher-repertoire-wiring.test.ts