research(agent-memory): mincut-gated forgetting — rejected, with evidence (ADR-345)

[ruvnet]

Summary

Nightly research run (2026-09-05, ADR-345): tests whether ruvector-mincut's existing RuVectorGraphAnalyzer can give ruvector-agent-memory's compaction a structural "don't evict the bridge memory" signal that the existing scalar CoherencePolicy cannot express.

Result: REJECT, with evidence. Two independent, measured failure modes:

1. Performance: RuVectorGraphAnalyzer::partition() scales from ~77ms (n=50 vertices) to ~11.4s (n=400), and cost ~1,800–2,700x the scalar baseline even at a deliberately tiny 84-memory corpus (threshold was 100x).
2. Non-determinism: on a hand-built graph with a provably unique weakest link, 30 repeated partition() calls on byte-identical input returned an empty/unusable result in 15/30 (50%) of calls.
3. At the corpus size the performance wall forced (84 memories), the structural bonus made zero measurable difference to bridge-memory survival vs. baseline (+0.0pp gap; threshold was ≥15pp).

A separate, independent piece shipped as a working capability regardless: compact_witnessed + EvictionWitnessChain, which closes a real gap — ruvector-agent-memory witnesses memory admission and (via a sibling crate) retrieval, but had nothing witnessing deletion. This reuses the crate's existing ADR-134 witness-record machinery verbatim (one new action_kind tag) and passed its own acceptance test cleanly (20/20 tamper-detection trials).

What changed

• crates/ruvector-agent-memory/src/graph_forget.rs (new, feature-gated behind mincut-forget, off by default): MincutGatedForgetting compaction policy (Soft/Hard variants).
• crates/ruvector-agent-memory/src/witnessed_compaction.rs (new, always on): compact_witnessed + EvictionWitnessChain.
• crates/ruvector-agent-memory/src/compaction.rs: factored CoherencePolicy's scoring into reusable weighted_importance().
• crates/ruvector-agent-memory/src/ops.rs: added action_kind::LEDGER_COMPACT_EVICT = 0xA7.
• crates/ruvector-agent-memory/examples/: mincut_gated_forgetting_bench.rs (the acceptance benchmark), mincut_scaling_probe.rs, mincut_determinism_probe.rs (the two feasibility/characterization probes referenced above — runnable, not one-off output).
• docs/adr/ADR-345-mincut-gated-forgetting.md
• docs/research/nightly/2026-09-05-mincut-gated-forgetting/README.md and gist.md

No existing behavior changes: CoherencePolicy/LruPolicy/LfuPolicy and compact() are untouched. graph_forget is opt-in and off by default.

(Note: this ADR was originally numbered 341, which collided with an existing ADR merged from another PR (#933) while this one was open. Renumbered to ADR-345 and rebased onto the current main in a follow-up commit on this branch.)

Benchmark command

cargo run --release -p ruvector-agent-memory \
  --example mincut_gated_forgetting_bench --features mincut-forget

Real benchmark results

Policy                           Bridge Surv.    Recall@10  Compaction (us)
----------------------------------------------------------------------------
CoherenceWeighted                       66.7%       100.0%               47
MincutGatedForgetting-Soft              66.7%       100.0%            86188
MincutGatedForgetting-Hard              66.7%       100.0%            85658

Tamper-detection trials (eviction witness chain)
  Detected 20/20 single-byte-flip tampers

Acceptance test
  Soft bridge-survival gap  (+0.0pp) >= 15pp : FAIL
  Hard bridge-survival gap  (+0.0pp) >= 15pp : FAIL
  Soft |recall delta| (0.00pp) <= 2pp                 : PASS
  Hard |recall delta| (0.00pp) <= 2pp                 : PASS
  Soft compaction slowdown  (1833.8x) <= 100x                : FAIL
  Hard compaction slowdown  (1822.5x) <= 100x                : FAIL
  Tamper detection (20/20)                          : PASS

=> REJECT: one or more mandatory acceptance thresholds failed (see above).

Full raw output, the scaling table (n=19..400), and the determinism-probe numbers are in the nightly README's "Benchmark Results" and "Failure modes" sections.

Acceptance result

REJECT (for production use as designed). graph_forget is retained, feature-gated and unpromoted, as a working reference implementation and as negative evidence so a future attempt doesn't rediscover the same two failure modes from scratch. witnessed_compaction is not gated by this rejection and ships as a usable capability today.

Darwin / Flywheel result

Neither tool exists in this repository: npx ruvector harness doctor --json fails (no such CLI), and no harness/darwin/flywheel subcommand exists anywhere in crates/ruvector-cli. npx metaharness --help resolves to a generic project-scaffolding tool unrelated to orchestrating research inside this repo. This capability-discovery result is recorded in the nightly README rather than assumed; the role-separation this process calls for (planner/researcher/engineer/critic) was performed serially in one session instead, with the design-probe → hypothesis-lock → run → analyze sequence documented in the README standing in for it.

Security review

No new cryptographic primitive. compact_witnessed reuses the crate's existing keyless-FNV-1a witness chain and its documented tamper-evidence scope unchanged (naive-edit detection, not adversary-resistant — see src/ops.rs's module docs). The mincut structural signal is advisory-only: a wrong or empty boundary set degrades to the already-reviewed CoherencePolicy behavior, never to an unsafe state.

Main limitations

• The effectiveness benchmark ran at only one corpus size (84 memories) — the size the measured performance wall forced. Whether the "no measured benefit" result holds at a larger, faster-computed scale is untested.
• The non-determinism reproduction uses one hand-built 19-vertex topology; its prevalence on larger, realistic clustered data is not independently isolated.
• ruvector-mincut's lower-level APIs (DynamicMinCut, ClusterHierarchy), which might avoid the measured overhead, were not tried — flagged as the primary next-research direction.

Production recommendation

Do not use MincutGatedForgetting in production. Do use compact_witnessed for any existing compaction policy that needs an auditable eviction trail — it has no mincut dependency and is enabled by default. Next research: repeat this exact, unmodified benchmark against ruvector-mincut's lower-level API surface once/if its performance and non-determinism are addressed upstream.

---

🤖 Generated with claude-flow

https://claude.ai/code/session_01TKszxZVDnLu1fCH2BAwFi5

[ruvnet]

CI status: Tests (core-platform) failed on this PR's head commit (d0071c2), but the failure is not in code this PR touches.

The job log shows 623/624 tests passed; the sole failure is ruvector-replication::replica::tests::test_promotion timing out at the nextest per-test limit (600.008s):

TIMEOUT [ 600.008s] (624/624) ruvector-replication replica::tests::test_promotion

That test (crates/ruvector-replication/src/replica.rs) is a trivial, synchronous, no-I/O unit test (three in-memory ReplicaSet operations, no sleeps, no network) that this PR does not modify — this PR only touches crates/ruvector-agent-memory. A test with that shape does not organically take 10 minutes; this looks like CI-runner-level contention (e.g. a stuck/hung neighboring test in the same nextest binary, or resource starvation) rather than a regression introduced here.

Re-running the failed jobs once to confirm this is transient, per the drive-to-green flake policy.

---

Generated by Claude Code