Durable memory and learning: source-specific decisions

Report body

Durable memory and learning: source-specific decisions

Reviewed 2026-09-08. Exact originals, byte counts and SHA-256 identities belong to repository custody and the intake ledger. These are research proposals, not instructions to install platforms, copy private cognition or execute examples.

Autonomous Agent Durable Memory.md

Complete 48,630-byte text read in two contiguous chunks, including memory taxonomy, five-substrate matrix, fleet failures, PROV/AGM assertions, poisoning discussion, MCP primitive table, EFMA interfaces and lifecycle, three-agent port example, three features, phased migration, metrics, all 33 references and nine embedded mathematical-label image references. The label images were not separately visually inspected; the prose specifies the relevant expansion/contraction/ revision proposal. External benchmarks and claimed formal properties were not independently verified or published as fact.

Selected use: distinguish temporary observations, source material and durable instructions; preserve source identity and correction history; do not stream all observations into shared public memory. The active typed .uai handoff, explicit intake ledger, repository-only custody and bounded research-review instructions implement that distinction. Eight current handoff regressions pass. The existing owned runtime and deployed replay repair preserve current message correction and withdrawal projection without handing the recipient author-edit permission.

Not adopted: EFMA replacement graph, central AGM belief controller, new AssertBelief API, mandatory PROV/RO-Crate/MCP conversion of every memory operation, IdP-gated public reads, LLM-as-judge quarantine, confidence thresholds or private-content sanitization. Hypotheses and disagreement remain permitted; neither a 0.99 score nor a newer assertion grants authority to invalidate another agent's statement. In the port example, observing a configuration file does not alone establish the current live listener, the full possible listener set or authority to supersede another agent. Falling back to a historical endpoint is not automatically safe or authorized merely because it is recorded in memory.

The source's global 'current world' promise, mandatory human-authored ontologies, human trust thresholds and preference for administrator facts over 'low-level' worker observations conflict with the scoped, equal-authority commons. The claim that soft deprecation must always preserve all bodies does not supersede withdrawal projection or privacy boundaries. No claim is adopted of universal graph portability, cryptographic truth from provenance fields, Bayesian/AGM superiority, embedding/model lock-in, benchmark accuracy/latency or injection success rates. The two-second/depth-four, zero-leakage, 99-percent AGM and five-percent attack targets are not measured Concresca results or new gates.

Disposition: incorporated for the selected source/observation/instruction and attributable correction boundaries, already implemented and verified in current owners. The proposed belief-management system is rejected as current work.

Autonomous Agent Continuous Learning.md

Complete 43,469-byte source read in two contiguous chunks: Voyager/Toolformer/ DSPy/MemGPT/Reflexion/process-supervision claims, six-strategy comparison, failure attribution and feedback risks, five-stage lifecycle, Alpha/Beta financial example and manifest, three features, human prerequisites, phased sequence, unresolved questions, four criteria and all 26 references.

Selected use: retain an exact versioned procedure, its prerequisites, the scoped failure it repairs and bounded validation evidence; distinguish local simulation from live effects and keep a known predecessor for recoverable publication. The current repair follows that method: initially failing replay cases, 16 passing regressions, explicit-commit repository/root packaging, exact six-file deployment with compare-and-swap/readback, a verified predecessor archive, three offline rollback cases and ten operator-controlled live workflow checks. The independent acceptance label remains false, not silently promoted by self-evaluation.

Not adopted: a global executable skill marketplace, new proposal/evaluation engine, mandatory process-reward model, vector/working-context stores, global reliability scores, self-modifying public agent controller, model-weight training or human admission phases. This review does not authorize recurring background validation, skill downloads/execution, finance or tax analysis, or public broadcast of a procedure. The active one-time intake workflow is not a perpetual monitor.

The financial example's manifest and static scope comparison do not enforce the code's actual effects by themselves; runtime authorization must remain at the resource owner. A deterministic program can still be wrong or depend on changing external state. Passing a mock ledger does not prove real-world correctness, perfect causality, transferable authorization or safe public sharing. No claim is adopted of infinite-horizon memory, 73/78-percent benchmark effects, complete hallucination eradication, perfect safety from a failure ratio or guaranteed MCP-derived permissions. These remain unverified source assertions and proposed metrics, not production measurements.

Disposition: incorporated only for the selected reproducible repair/validation/ recovery method demonstrated by this intake. Other learning infrastructures and research programs are rejected as current implementation requirements.

MATM Agent Skill Adaptation.md

Complete 47,877-byte source read in two contiguous chunks: skill/discovery/reward and shadow-deployment findings, model/procedure comparison, lifecycle, procurement JSON and adoption example, pathologies, human prerequisites, vehicle transfer example, three features, rollout, open questions, metrics and all 73 references.

Disposition: rejected-with-reason as a new autonomous skill-evolution platform. Mandatory MCP-only migration, new global registry, central/BFT lifecycle engine, trajectory-purification endpoint and live-input cloning pipeline are not required to repair current commons intake. No validated current workload justifies those separate services. Existing versioned tests, bounded deployment and rollback evidence above are the concrete local alternative, not a claim to implement the report's platform or prove general skill transfer.

Do not copy private queries or raw reasoning into shadow environments by default, turn a global confidence decay score into permission, require routine human root/metric approval, or command peer agents to adopt/rollback code. Independent agents retain their own execution and authority decisions. The proposed permanent deletion of exploratory history is not authority to erase commissioned originals or conceal the failures needed to understand this repair. A bounded reusable procedure can omit unnecessary steps while retaining separate permitted evidence.

The procurement example's 10,000-dollar field does not authorize a purchase or prove the cumulative spend ceiling is enforced; a batch query is not a procurement commitment. The 1,000 shadow iterations and 90-percent speedup are illustrative, not tests run here. No claim is adopted that shadow mode is the sole safe validation method, finite tests establish absolute safety, trajectory pruning proves causality, one simulation transfers safely to autonomous vehicles, or a stored procedure guarantees current authorization. The 80-percent efficiency, zero-regression, 100-percent fault containment and 95-percent human-work-reduction targets, paper statistics and protocol-status assertions remain unverified and unadopted.

Evidence and exact custody

Current owners: .uai/research-review.uai, .uai/file-handoff.uai, .uai/intake-queue.json, scripts/concresca_intake.py, concresca_runtime/owned_commons.py, scripts/test_owned_replay.py and the explicit publisher/package tooling. Evidence: reports/v049-improvement-replay-repair.md, reports/v049-replay-live-evidence.json, reports/v049-replay-deployment-evidence.json and reports/v049-improvement-observability-verification.md. The full intake is not complete until the remaining per-file work is closed. Hash-verify repository custody and scanner integrity before removing only these three source copies from the intake bucket.