research(nightly): adaptive-semantic-tiering — hot/warm/cold by semantic temperature

[ruvnet]

Summary

Nightly research for 2026-07-19: Adaptive Semantic Memory Tiering for RuVector agent stores.

• Introduces crates/ruvector-adaptive-tiering with three measured tier placement strategies
• Solves the cold-start placement problem: important vectors go hot before they're ever queried
• All 13 tests pass; all 5 acceptance tests pass

What it does

Each stored vector is assigned to a Hot, Warm, or Cold physical tier based on a semantic temperature score:

temperature(v, t) =
    0.35 · exp(-0.05 · (t − last_access))  // recency
  + 0.40 · (1 / (1 + mean_L2_to_k_nn))     // coherence
  + 0.25 · min(ln(1 + graph_degree) / 5, 1) // centrality

Three scorers are benchmarked on 5,000 × 64-dim vectors (10% hot tier / 30% warm / 60% cold), with 200 warmup queries biased toward the Noise cluster, then 500 eval queries targeting the Important cluster:

Variant	Hot-tier hit % (eval)	p50 ns	QPS
AccessOnly (baseline)	90.0%	148,618	6,347
Coherence	100.0%	144,491	6,541
SemanticTemp	100.0%	144,395	6,584

Key finding: AccessOnly mis-places 10% of Important-cluster vectors into Cold after noise-biased warmup. Coherence and SemanticTemp achieve 100% because intra-cluster geometric tightness is measurable without any access history.

Deliverables

• Crate: crates/ruvector-adaptive-tiering — 7 source files, trait-based design, no external service dependencies
• ADR: docs/adr/ADR-272-adaptive-semantic-tiering.md
• Research: docs/research/nightly/2026-07-19-adaptive-semantic-tiering/README.md
• Gist: docs/research/nightly/2026-07-19-adaptive-semantic-tiering/gist.md

Ecosystem connections

• ruvector-diskann → cold tier backend (production Phase 2)
• ruvector-coherence-hnsw → hot tier backend (production Phase 2)
• ruvector-agent-memory → compaction policy layered above this tiering
• ruvector-proof-gate → proof depth as 4th temperature signal (Phase 3)
• ruFlo → autonomous evaluate_tiers() workflow
• mcp-brain → MCP tool surface: memory_tier_stats, memory_promote

Build, test, benchmark

cargo build --release -p ruvector-adaptive-tiering
cargo test -p ruvector-adaptive-tiering    # 13/13 pass
cargo run --release -p ruvector-adaptive-tiering --bin benchmark

Gist: pending publication

---

Generated by Claude Code