Stash profile verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

I checked the memory profile, including every axis, unit assessment, finding basis and coverage note, against the handed runtime, memory, epistemic and workflow members. I applied the supplied reconciliation's supersessions: memory object declarations and the memory pattern/failure routes are the canonical referents. The narrower epistemic routes remain separate witnesses where their referents differ from the combined memory routes. No target source was read; this judgment adds no source evidence or quotations.

Storage, form and lineage

The storage inventory accounts for content, access/control metadata, vectors and cache. PostgreSQL supports rdbms; pgvector supports vector. Relational edge rows do not independently support graph storage. Grouping the relational parts does not conceal a different recorded substrate. Provider weights and static products are explicitly outside the use-accumulated memory scope, not classified through their readable interfaces. See MEM-OBJ-memory-vectors, MEM-OBJ-embedding-cache, MEM-OBJ-settings and RT-CMP-openai-embedder.

The form units distinguish interpreted text, fixed-rule fields and edges, and dense numerical representations. The separate language-consumption unit for relationship words and fact fields prevents treating their symbolic matching role as their only form. The known assessment concerns recorded encodings, not complete knowledge of consumers or original authors. See MEM-OBJ-fact, MEM-OBJ-relationship and MEM-BAP-maintenance-input.

Lineage correctly separates import at the Stash boundary from internal derivation. Failure-review experience supports local trace-extracted lineage; arbitrary episode assertions leave the broader trace-versus-other-compilation split partial. Known embedding production does not purport to resolve upstream assertion provenance. See RT-RTE-remember-write, MEM-OBJ-episode and MEM-RTE-failure-derivation.

Authority, admission and curation

Authority findings identify retained parts at their respective consumers: client reference, afforded standing guidance, maintenance evidence, experience-fed durable updates, numerical/priority ranking, scope/window routing, enforced exclusions and retained transition/dimension criteria. They do not infer client obedience from delivery. Pattern consumers and further settings uses remain unresolved, so the axis does not claim complete authority coverage. See MEM-BAP-retained-advice, MEM-BAP-maintenance-input, MEM-OBJ-consolidation-progress, MEM-RTE-goal-maintenance and EPI-RTE-hypothesis-control.

Write agency is assessed at admission rather than authorship or SQL execution. The automatic maintenance/cache/bookkeeping findings do not upgrade caller-controlled alternatives whose human approval is unknown. Starting maintenance is not approval of each generated result. See RT-RTE-consolidation-run and RT-RTE-context-handoff.

Curation findings distinguish episode compression, similarity-based duplicate suppression, entry revision, promotion, confidence decay and selector-effective invalidation. The profile does not mistake the anti-copy gate for semantic validation or validity metadata for complete withdrawal. Hypothesis content refinement is explicitly afforded, not wired or observed; its note preserves the timestamp defect. Other wired entry updates independently support evolve. Possible semantic novelty remains not determinable, rather than being inferred from synthesis terminology. See EPI-RTE-fact-synthesis, MEM-RTE-edge-extraction, MEM-RTE-hypothesis-maintenance and MEM-RTE-fact-withdrawal.

Read-back and trace-fed updates

Direction units distinguish consumer-requested recall, typed/focus reload and operational lookups from automatic assembly of retained evidence for maintenance models. A foreground maintenance request does not turn the reasoner's supplied evidence into a model-requested pull. Conversely, requested tool-result serialization is not a second push. Pattern read-back stays afforded and unresolved getter consumers are not silently omitted. See RT-RTE-recall-read, MEM-RTE-typed-read and MEM-RTE-pattern-extraction.

Push signals have actual selection witnesses: coarse namespace/checkpoint/status windows, embedding-selected episode clusters, and identity-matched old fact values delivered to the contradiction comparator. The identifier finding is not inferred merely from IDs attached to results. Requested recall's query embedding is not classified as a push signal. Unestablished alternative consumers prevent a complete signal inventory. See MEM-OBJ-consolidation-progress, RT-RTE-consolidation-run and EPI-RTE-contradiction-review.

Trace-learning positives retain automatic write, durable result and later-consumer chains. Failure-derived episodes/facts supply the strongest local witness; session-bearing episode consolidation is qualified because arbitrary assertions are also admitted. Neither authentication nor improved capacity is added as a membership condition. The original report/summary inputs do not establish recorded turns, tool records, environmental event streams or linked trajectories. Therefore the trace-source axis properly remains not determinable, not absent or inapplicable. See MEM-RTE-failure-derivation, MEM-OBJ-episode and RT-RTE-consolidation-run.

Conflicts and disposition

The theory-builder condition-3 label disagreement does not determine any emitted memory value: neither position supplies a criticism instance, and the profile claims no exercised theory-building or improvement. See MEM-RTE-hypothesis-maintenance and EPI-RTE-hypothesis-assessment. The workflow account also preserves the distinction between implemented admission/delivery and observed successful use.

The grouped units are coarse but retain the consequential differences in form, admission, consumer, transformation and evidence basis. Their unresolved alternatives are explicit. I found no unsupported emitted comparison value, unjustified complete assessment or concealed material branch. The profile's own partial and not-determinable assessments already express the remaining uncertainty; those uncertainties are not additional verification limits.

Blockers

none

Limits

none