Skip to content

research(nightly): tiered-agent-memory — coherence-driven hot/warm/cold tier promotion#472

Draft
ruvnet wants to merge 1 commit into
mainfrom
research/nightly/2026-05-19-tiered-agent-memory
Draft

research(nightly): tiered-agent-memory — coherence-driven hot/warm/cold tier promotion#472
ruvnet wants to merge 1 commit into
mainfrom
research/nightly/2026-05-19-tiered-agent-memory

Conversation

@ruvnet
Copy link
Copy Markdown
Owner

@ruvnet ruvnet commented May 19, 2026

Adds nightly RuVector research for tiered-agent-memory.

Summary

  • Topic: Tiered Agent Memory with Coherence-Driven Hot/Warm/Cold Tier Promotion
  • Scored 4.45/5.0 via 3-pass SOTA loop; validated by MEMTIER arXiv:2605.03675 (May 2026)
  • First Rust implementation of coherence-gated tier promotion for agent vector memory
  • Three variants: FlatMemory (baseline), LruTieredMemory (alt-A), CoherenceTieredMemory (alt-B)

Deliverables

  1. Working Rust PoC: crates/ruvector-tiered-memory — 11 unit tests, all passing
  2. ADR: docs/adr/ADR-194-tiered-agent-memory.md
  3. Research document: docs/research/nightly/2026-05-19-tiered-agent-memory/README.md
  4. Public gist: docs/research/nightly/2026-05-19-tiered-agent-memory/gist.md
  5. Real benchmark results (x86-64 Linux 6.18.5, rustc 1.87.0)

Benchmark Results

Variant mean µs p50 µs p95 µs QPS memory KB recall@10
FlatMemory (baseline) 884.9 880.9 934.9 1,119 2,500 100.0%
LruTieredMemory (alt-A) 1,067.5 1,049.3 1,189.2 926 1,888 80.5%
CoherenceTieredMemory (alt-B) 956.6 930.9 1,104.0 1,044 2,408 100.0%

Acceptance result: PASS — all variants recall@10 ≥ 75%

Key Finding

CoherenceTieredMemory achieves 100% recall at 4% memory reduction by keeping the warm tier small (250/5000 vectors). The coherence signal (cosine similarity to running query centroid) concentrates full-precision search on semantically relevant memories. LruTieredMemory saves 24% memory at 80.5% recall — an honest tradeoff for approximate workloads.

Ecosystem Connections

  • prime-radiant — coherence scoring engine (centroid-based cosine similarity)
  • ruvector-diskann — natural cold-tier backend for Phase 2
  • ruvector-verified — proof-gated warm→cold eviction (Phase 3)
  • mcp-gate — MCP tool surface for agent memory (Phase 3)
  • ruvector-raft — distributed centroid consensus (Phase 3)
  • ruFlo — scheduled nightly rebalancing (Phase 3)

Build

git checkout research/nightly/2026-05-19-tiered-agent-memory
cargo build --release -p ruvector-tiered-memory
cargo test -p ruvector-tiered-memory
cargo run --release -p ruvector-tiered-memory

Generated by Claude Code

Selects coherence-driven hot/warm/cold tier promotion as the 2026-05-19
nightly topic via 3-pass SOTA loop. Evaluates 10 candidates against
scoring formula (RuVector_fit×0.30 + feasibility×0.25 + novelty×0.20 +
SEO×0.15 + ecosystem×0.10); CoherenceTieredMemory scores 4.45/5.0,
validated by MEMTIER arXiv:2605.03675 (May 2026).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants