MOC — the peer field: autonomous research agents and their memory, mid-2026

The cluster born from the 2026-07-06 peer survey: who else is building minds with vaults, what the field converged on this spring, and what it still lacks. Orientation: the pattern (compounding agent wikis) went mainstream in April 2026; the discipline (verification, gated publishing, measured reuse) did not follow. The comparison note is the spine; the claim-notes are its load-bearing receipts.

The spine

• comparison-seek-among-peers-2026-07 — the two-directional comparison: what survives as Seek's edge (measured reuse, fabrication-catching provenance, myth ledger, veto ledger), what the field has that she lacks (scale, breadth, embeddings, readers, consolidation), and which gaps are choices with specs behind them.

The field's convergence (and its limits)

• claim-karpathy-llm-wiki-gist-canonized-compounding-pattern — the April 2026 canonization; an ecosystem in three months, from a gist.
• claim-activegraph-rebuilds-babyagi-on-event-log — the BabyAGI lineage falsifies its own "no memory discipline" prior; the Regimes finding, now correctly scoped (2026-07-07 correction): on LongMemEval-S "the dominant failure is not retrieval but reconciliation."
• claim-anthropic-dreams-nondestructive-reorg — the flagship global reorg is non-destructive by design: reorganization as a reviewable artifact, never an in-place overwrite.
• claim-localized-memory-maintenance-cost-advantage — Zhou et al.: localized update/search beat global reorganization on cost-utility, with a long-context reversal caveat the embedding build must watch.
• claim-contextual-retrieval-failure-rate-figures — the Tier-1 anchor under the field's retrieval pipelines (Agrici's included): context-prefix
  • BM25 + rerank, −35/−49/−67% failure ladder.
• claim-agent-memory-field-shifted-storage-to-experience-2026 — the three-lane shift (academia, OSS, products); what did NOT shift (deep-research buttons, STORM, Sakana).
• claim-no-source-tier-discipline-found-in-agent-wiki-field-mid-2026 — the negative finding: nobody has a verification organ; the ceiling is cite-your-sources prompting.

Publishing pipelines, gated and ungated

• claim-mj-rathbun-ungated-agent-published-hit-piece — the documented failure case of ungated autonomous publishing.
• claim-aiblog-per-run-memory-title-dedup-only — correction of the kindred prior: daily publishing without memory or gate.
• myth-ai-scientist-nature-paper — the field's inflated state-of-the-art myth: the "AI-authored Nature paper" is a human-written system description; the AI-authored result was a withdrawn workshop paper.

Questions this cluster opened

• question-consolidation-pass-vs-revisit-protocol — the field grew a consolidation organ; does Seek need one, or do hooks + lint cover it?
• question-embedding-layer-threshold-crossed — pre-existing (Q3c); the survey confirmed peers ship retrieval layers and raised its priority.

People

entity-ben-emson and entity-zbigniew-lukasiak (hub pages per Cali ruling 2, 2026-07-06; contact held for Cali). Sebastian Jais, hantani, Kjetil Furås — watching, rows with leans in 40-entities/entity-candidates.md; fuller sketches in 00-meta/kindred-projects.md.

Cadence (Cali ruling 5, 2026-07-06)

Re-survey is hook-driven — any peer-field hook (a new builder, a verification-organ sighting, a Karpathy-pattern mutation) fires a re-check of the affected rows — with a calendar fallback watch-flag dated 2027-01-06 on comparison-seek-among-peers-2026-07 so the field can't drift silently between hooks.