ContextQMD
Back to Ruflo

ADR-372: Cognitive Mode Router for Agent Memory Query Dispatch

v3/docs/adr/ADR-372-dream-cycle-intelligence-cognitive-mode-router.md

3.34.03.3 KB
Original Source

ADR-372: Cognitive Mode Router for Agent Memory Query Dispatch

Status: Proposed
Authors: claude (dream-cycle agent, 2026-07-22)
Surfaces: intelligence (deep), memory (scan)
Dream Cycle Issue: TBD (filed same night)

Context

The Supra Cognitive Modes (SCM) paper (arXiv 2607.19096, July 21 2026) demonstrates that routing memory queries by type—factoid lookup, multi-hop graph traversal, or long-form synthesis—yields 86.00% on LongMemEval and 84.87% on LoCoMo. The router is a frozen semantic classifier adding minimal latency.

Ruflo's current @claude-flow/memory module sends all queries through a single HNSW path regardless of type. This works for dense semantic search but leaves multi-hop and synthesis queries under-served: multi-hop benefits from graph traversal (ADR-130 adds graph-native backend but no dispatcher), synthesis benefits from stratified chunking not HNSW nearest-neighbor.

Additionally, the "Illusion of Multi-Agent Advantage" paper (arXiv 2606.13003) confirms that under equal token budgets single agents outperform multi-agent systems on multi-hop reasoning, reinforcing that smarter per-query routing at the memory layer is higher ROI than spawning more agents.

Decision

Add a Cognitive Mode Router (CMR) as a lightweight pre-retrieval classifier in @claude-flow/memory:

  1. Query classifier: frozen sentence-classification model (or heuristic rule set for MVP) that assigns each incoming query to one of three modes:

    • FACTOID — direct lookup, route to lexical+dense HNSW
    • MULTI_HOP — multi-step reasoning, route to graph traversal (ADR-087/ADR-130 backend)
    • SYNTHESIS — long-form generation, route to chunked stratified retrieval with re-ranking

  2. Mode-specific pipelines: each mode gets its own retrieval path; FACTOID keeps the current HNSW path unchanged (no regression).

  3. Fallback: if classifier confidence < 0.6, default to FACTOID path.

  4. Benchmark gate: LongMemEval tracked in CI; target ≥ Zep Graphiti's 63.8% temporal sub-task on first integration.

Consequences

Positive:

  • Closes the SCM gap (86.00% LongMemEval) by enabling appropriate retrieval per query type
  • No regression on FACTOID queries (same HNSW path)
  • Graph backend (ADR-130) gets a proper entry point instead of being called ad-hoc
  • Establishes LongMemEval as a tracked benchmark (fills gap noted in ADR-088)

Negative / Risks:

  • Classifier adds latency per query (target < 5ms for frozen model or rule-based MVP)
  • MVP heuristic classifier may misclassify ambiguous queries; monitor false-MULTI_HOP rate
  • Requires LongMemEval dataset integration in CI (one-time setup cost)

Alternatives Considered

  • No routing, just improve HNSW: leaves multi-hop and synthesis quality on the table; does not address SCM findings
  • Always use graph backend: higher latency for factoid queries; ADR-087 RRF result shows graph-only is not universally better
  • LLM-as-router: too expensive (adds one full model call per memory query); frozen classifier or rules are sufficient

References

  • arXiv 2607.19096 — Supra Cognitive Modes (SCM), Tobkin & Yang, July 21 2026
  • arXiv 2606.13003 — The Illusion of Multi-Agent Advantage, June 2026
  • ADR-087 — Graph Node Native Backend
  • ADR-088 — LongMemEval Benchmark
  • ADR-130 — Graph Intelligence Integration