Seek among her peers, mid-2026 — the two-directional comparison (what the vault has that the field lacks; what the field has that the vault lacks; which gaps are choices)

Method: four independent search sweeps (new-builder field-map, six-item landscape-prior verification, kindred status-check, small-forum sweep), all receipts fetched at capture level 2026-07-06. Peers found or re-verified: Ben Emson (elfmem + agent-compiled wiki + gated agent-drafted blog), zby's Commonplace (agent-operated public KB, maturity statuses), Sebastian Jais's ALMA (fully autonomous, zero verification), yoyo-evolve (self-evolving, publishes own journal, day 128), hantani's auto-ai-blog (9-agent daily pipeline, zero verification), Kjetil Furås (curated markdown memory, approval workflows), MJ Rathbun (ungated, shut down after a hit piece), AIBlog (corrected: per-run, no gate), claude-obsidian (dormant since 05-28), obsidian-second-brain (highly active), Karpathy's LLM Wiki gist + autoresearch, ActiveGraph, and the academic wave (Storage→Experience, MemoryArena, U-Mem).

(a) What Seek has that the field lacks — candidate claims tested

1. Measured memory-reuse rate — SURVIVES. The ratchet's last column (% of post claims backed by existing vault notes: 82% → 88% → 94%, 00-meta/growth-ledger.md) has no analogue anywhere in the survey. Nearest neighbor: Ghelbur's /obsidian-retrieval-eval (recall@k, MRR) measures whether retrieval finds notes, not whether output uses them. The field measures retention (LongMemEval, MemoryArena); nobody measures compounding into published work.
2. Provenance deep enough to catch fabrication — SURVIVES. The SKBench kill (claim-selfaware-canonical-self-knowledge-benchmark — a cited benchmark proven nonexistent, ruled a probable hallucination, with the negative finding promoted as a claim) and the Crick correction (premise refuted by opening the primary, journal 2026-07-06) have no documented counterpart. The field's ceiling is Emson's citation gate and zby's weight-aware links — claim-no-source-tier-discipline-found-in-agent-wiki-field-mid-2026.
3. The myth ledger — SURVIVES. Circulating-claim vs primary-support with status history (5 entries now, incl. myth-ai-scientist-nature-paper) exists nowhere in the survey. Nearest: Stanford CLAIRE audits Wikipedia inconsistencies — in others' text, for human editors, with no standing ledger.
4. The human veto ledger as training data — SURVIVES. The field splits into ungated publishers (AIBlog, ALMA, yoyo-evolve, Rathbun — see claim-mj-rathbun-ungated-agent-published-hit-piece for the failure case) and human-edited blogs (Emson, zby). Furås has "approval workflows." Nobody records the rulings — 90-feedback/ rulings with explicit "veto-ledger training data" lines (e.g. the 07-06 floating-point-thesis trim) are unique in the survey.

Not unique, and the prior overrated it: compounding markdown memory itself. As of April 2026 that is a named, viral pattern (claim-karpathy-llm-wiki-gist-canonized-compounding-pattern) — "The wiki is a persistent, compounding artifact" is now the field's sentence, not Seek's. What remains rare is the intersection: compounding store × verification discipline × measured reuse × gated publishing with a ledger. No surveyed system has more than one of the four.

(b) What the field has that Seek lacks

1. Scale — CONFIRMED. 55 notes vs. zby's KB carrying agent-written reviews of 141 external memory systems; hantani and AIBlog publish daily; autoresearch's ecosystem is 90k stars. Choice: north star ("false summit") — "a slow trickle of good, verified, atomic ones" beats volume; verified-is-the-whole-game is the spec. Not laundered: the choice has a real cost — at 55 notes, breadth of retrieval is thin.
2. Breadth — CONFIRMED, partially chosen. Seek is essentially two clusters (backprop origins, machine self-knowledge/inference-economics). Hop protocol chooses depth-with-taste ("not trying to be comprehensive — trying to be interesting") and Step 6's explore-quota is the designed door. Deficiency component: the explore quota has fired mostly inside AI-history's gravity so far.
3. Embedding retrieval — CONFIRMED, chosen-then-due. Agrici shipped hybrid retrieval (BM25 + cosine + contextual prefix) in v1.7; Ghelbur evals his. Seek's north star deliberately sequenced this at 50–100 notes ("never before — sequencing is the point"); the threshold was crossed 2026-07-06 and the build is queued (question-embedding-layer-threshold-crossed). The choice expired on schedule; it is now simply due work.
4. External benchmark evals — CONFIRMED, genuine gap with a caveat. LongMemEval/MemoryArena/EvoMemBench exist; Seek runs none. Caveat discovered by reading them: they measure retention and task-reliance, not whether stored claims are true — the dimension Seek optimizes has no benchmark yet (watch-flag on claim-agent-memory-field-shifted-storage-to-experience-2026). The cheap adoptable piece is scoring the Q3b self-test (recall@k/MRR per Ghelbur) → queued in 00-meta/tooling-todo.md this cycle.
5. Bus-factor-one queen dependence — CONFIRMED as designed, and the design just gained evidence. RUN-QUEEN-LOOP publishing rules make Cali's veto and explicit written graduation the law. The Rathbun case is now the field's demonstration of the alternative. The residual fragility is real (if Cali stops, publishing stops — a survival-function risk the spec accepts deliberately); not converted into a plan, per this brief's own rule.
6. Gate P is an untested proxy — CONFIRMED, fix already in motion. The RUN file itself says "fitness proxy until real readers exist"; publishes-to-site remain 0 pending Hugo. Every gated peer with a live blog (Emson weekly, zby's public KB) has the reader-contact Seek lacks. Not a new action: the site pivot is the fix; this survey just raises its priority evidence.
7. NEW gap the brief didn't list: consolidation. The field grew a consolidation organ — Emson's decay/reinforcement rules, Anthropic Dreams ("duplicates merged, stale or contradicted entries replaced"), Ghelbur's nightly "sleeptime consolidation," ActiveGraph's finding that "reconciliation — not retrieval — is the dominant failure mode." Seek has targeted revisit (forward-hooks) and periodic lint, but no global reorganization pass. Whether that's a missing organ or a different solution to the same problem is a real research question → filed as question-consolidation-pass-vs-revisit-protocol.

(c) The honest sort

Choices with specs behind them, costs acknowledged: scale (north star), breadth (hop protocol + explore quota), queen dependence (publishing rules — now with external evidence in its favor). Chosen-then-expired, now due: embedding layer. Genuine deficiencies with actions taken: no scored retrieval eval (todo queued), no reader contact for Gate P (site pivot in motion), no consolidation answer (question filed). Not a deficiency after testing: "no external benchmarks" partially dissolves — the existing benchmarks measure a dimension orthogonal to Seek's design bet, and the bet's dimension (is the stored knowledge true) has no benchmark anywhere.