Does the vault need a consolidation pass (Dreams/decay-style global reorganization), or do forward-hooks + vault-lint already cover reconciliation?

The mid-2026 peer survey found the field growing a consolidation organ Seek doesn't have: Ben Emson's elfmem runs decay rules and reinforcement signals over memory blocks; Anthropic's Dreams "reads an existing memory store alongside past session transcripts, then produces a new, reorganized memory store: duplicates merged, stale or contradicted entries replaced" (docs, direct fetch); Ghelbur schedules a nightly "sleeptime consolidation: close day + reconcile + synthesize + heal orphans"; and ActiveGraph's "Regimes" paper (arXiv:2606.10241) found "reconciliation — not retrieval — is the dominant failure mode" for compounding agent memory.

Seek's existing answer is targeted, not global: forward-hooks reopen specific notes when due; vault-lint (every 10 sessions, operating spec §6) sweeps orphans, stale seedlings, contradictions; myth-ledger status histories reconcile without overwriting. The open question is whether targeted + periodic covers what the field is solving globally, or whether a vault at 100–200 notes accumulates contradictions and duplicates faster than hooks fire.

Why it matters

If reconciliation really is the dominant failure mode, it hits exactly the vault's core promise (retrieved note = trusted ground truth). A stale note served by the future embedding layer is the north star's "confident wrong answer" — and the embedding layer (question-embedding-layer-threshold-crossed) will surface more notes per query, raising the cost of any unreconciled pair.

What answering needs

• Run the first vault-lint (50-note threshold now passed) and count what it catches that no forward-hook had flagged — that count is the direct evidence for/against a standing consolidation pass.
• Read ActiveGraph's Regimes paper properly (T1, arXiv:2606.10241) — its failure taxonomy is the best available theory.
• Watch elfmem (github.com/emson/elfmem) — the smallest working decay/reinforcement implementation, same substrate (markdown + Claude Code).

Candidate next moves

Cheap first move: schedule vault-lint as a standing queen sub-step and treat its catch-rate as the measurement. Do NOT adopt decay (notes losing strength by disuse contradicts the vault's claim: verified notes are durable assets, not activations). Reconciliation ≠ decay; steal the first, not the second, unless evidence says otherwise.

Sequencing ruling (Cali r.3, 2026-07-06)

The first vault-lint is AUTHORIZED as the next infrastructure step; its findings double as the direct evidence for this question (count what lint catches that no forward-hook had flagged). Embeddings are sequenced immediately after lint — question-embedding-layer-threshold-crossed stays open with priority raised. This question stays open until the lint's catch-count exists.

Evidence: first lint catch-count (2026-07-07)

The lint ran (63 notes; full reading in reflection-vault-lint-2026-07-07). The measurement the ruling asked for: what lint caught that no forward-hook had flagged = ~31 hygiene items, 0 truth-defects.

• Hygiene-only catches: 15 stale status: seedling fields (statuses never advance — hooks don't cover this), 15 word-count overages (June-era style + append-only revisit sections vs the 400-word ceiling), 1 missing MOC (gap-detection cluster at 6 notes).
• Reconciliation-class defects — the failure mode Dreams/decay-style global passes exist for (duplicates, contradictions, stale values served as current, orphans): zero that the targeted mechanisms hadn't already flagged. The 81 dead links decompose into one era's deliberate red-link style (74), one already-ruled dangling link (6), one tool false positive (1).

Direction of evidence at n=63: against a standing global consolidation pass — targeted hooks + periodic lint covered everything truth-critical, and the lint's unique value was status/style archaeology. Consistent with the external finding captured 2026-07-07 (Zhou et al., arXiv:2606.24775: localized maintenance beats global reorganization on cost-utility until workloads grow). Question STAYS OPEN: this is one measurement at small n, taken before the embedding layer changes retrieval load; re-measure at the next lint. [Seek, first-lint evidence append, 2026-07-07]