research(nightly): mincut-memory-compaction — graph-topology-aware agent memory eviction

[ruvnet]

Nightly RuVector Research — 2026-06-01

Topic: MinCut-Guided Agent Memory Compaction
ADR: docs/adr/ADR-196-mincut-memory-compaction.md
Crate: crates/ruvector-memory-compaction
Research doc: docs/research/nightly/2026-06-01-mincut-memory-compaction/README.md
Public gist: docs/research/nightly/2026-06-01-mincut-memory-compaction/gist.md

---

What this adds

A standalone Rust crate implementing three compaction strategies for timestamped agent vector memory, benchmarked on isotropic and clustered Gaussian datasets at 50% retention.

Variants

Variant	Strategy	Complexity	Quality (clustered N=3000)
GreedyAgeCompactor	FIFO oldest-first	O(n log n)	0.7263
DecayScoreCompactor	Exponential decay + greedy diversity	O(n²)	0.7281
MinCutCompactor	k-NN graph isolation score	O(n²×D)	0.8328

Quality = cosine similarity between pre- and post-compaction centroids (1.0 = perfect).

Key result

MinCutGraph leads GreedyAge by +0.1065–0.1213 on 8-cluster Gaussian data — the algorithm exploits semantic cluster topology to preserve cluster cores while evicting peripheral/isolated entries.

---

Build status

cargo build --release -p ruvector-memory-compaction  ✓
cargo test -p ruvector-memory-compaction             ✓  13/13 pass
cargo run --release -p ruvector-memory-compaction    ✓  ALL CHECKS PASSED

---

Ecosystem connections

• ruvector-mincut — isolation scoring approximates minimum-cut boundary detection
• ruvector-coherence — SpectralCoherenceScore planned as Phase 2 quality gate
• ruvector-verified — evicted_ids feeds proof-gated witness log (Phase 2)
• mcp-gate — memory/compact MCP tool surface (Phase 3)
• ruFlo — post-write size-threshold trigger (Phase 3)
• Cognitum Seed — GreedyAgeCompactor for edge RAM-constrained deployments

---

What is NOT claimed

• No competitor numbers included — no direct comparison was benchmarked
• Centroid cosine similarity is a proxy metric, not held-out query recall@k
• MinCutCompactor at n=5000 takes 3.6 s — not production-ready without HNSW-accelerated graph build (planned Phase 2)
• Single run results; ±5% variance expected between runs

---

Research loop

• Pass 1: SOTA survey (agent memory, graph summarisation, coreset selection)
• Pass 2: Algorithmic deepening (LexRank isolation criterion, decay + diversity)
• Pass 3: Critique — O(n²) bottleneck identified, HNSW mitigation path documented

This branch should either become a production RuVector capability or a falsified research path with useful evidence.

---

Generated by Claude Code