Memori profile verification

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

Verification

I checked all ten axes, their local findings and evidence bases, natural units, coverage assessments and comparison rationale against the supplied records, resolving superseded identities through reconciliation. I also checked the profile author's correction answer against the earlier objection and the current profile. No source reads, new records or new quotations were needed.

Storage and representational form distinguish SQLite rows, graph/vector access structures and benchmark files, and separate language, fixed-rule symbolic fields and dense numerical representations. Grouped inventory units retain the supporting part in each finding rather than assigning every value to every object. The hosted service interface and visible JSON envelope do not classify the private retained representation. These classifications are supported by MEM-OBJ-conversation-message, MEM-OBJ-knowledge-graph, MEM-OBJ-content-embedding, MEM-OBJ-benchmark-artifacts and MEM-OBJ-agent-memory.

Lineage keeps recorded interaction origins, client compilation and benchmark import distinct. Private extraction semantics, immutable vector-model provenance and native mock/fallback origins remain unresolved. Write agency follows the actual admission mechanism: software capture and mapped writes are automatic, while a generic caller or workflow initiation does not establish a human decision about retained content. Hosted initialization, feedback disposition and explicit withdrawal control are not silently classified as manual. See RT-RTE-augmentation-admission, MEM-OBJ-entity-fact, MEM-OBJ-fact-access, MEM-RTE-benchmark-reload, MEM-RTE-feedback-submission and MEM-RTE-entity-withdrawal.

Behavioral authority separates text supplying reference at inference, retained content and vectors affecting rank, and prior summaries feeding a durable update and selecting the next extraction window. Claimed standing-order use does not become enforcement or observed obedience. Curation distinguishes summary revision, read-time summary deduplication, recurrence-driven salience and destructive forgetting; hosted consolidation remains claimed rather than inheriting client wiring. These mappings follow MEM-OBJ-conversation-summary, MEM-OBJ-fact-access, MEM-RTE-fact-selection, MEM-RTE-entity-withdrawal and RT-BAP-host-memory-guidance. Uninspected independent graph/process consumers and report/feedback use prevent complete authority coverage, not the supported local effects.

Read-back direction separates backend requests from automatic selection and context supply within a chain. Requested host tool results remain pull; automatic SDK provider-context enrichment and benchmark answer-context assembly support push without implying unsolicited background messaging. SDK signal findings name query similarity, lexical scores, actual entity/conversation/fact matches and separate coarse gates. See RT-RTE-python-invocation, RT-RTE-typescript-hooks, MEM-RTE-fact-selection and MEM-OBJ-conversation-message.

The earlier benchmark identifier objection is resolved. The current benchmark push-signal unit no longer emits identifier, retains supported dense, lexical and coarse findings, and is partial rather than known. Its note and dependent rationale now distinguish requested per-conversation index access from automatic selection of delivered memories. This matches MEM-RTE-benchmark-reload and MEM-OBJ-benchmark-artifacts. The correction does not withdraw the independently supported SDK identifier findings or assert that an unestablished benchmark selector is absent.

Trace learning identifies the automatic interaction-fed write to durable facts/summaries and their later context consumer. Its wired classification does not require observed improvement or known server semantics. Host resume derivation and optional tool-trace consumption retain claimed bases and partial coverage. Trace source follows conversational and tool activity rather than hook labels; no event-stream or trajectory finding is inferred merely from triggering or ordering. See RT-RTE-augmentation-admission, MEM-OBJ-conversation-summary, MEM-RTE-fact-selection, MEM-OBJ-agent-memory and RT-RTE-hermes-provider. Native alternatives, upstream benchmark ancestry and unresolved feedback or independent graph/process consumers remain uncertain rather than becoming system-wide negatives.

All ten axes remain partial, with named opaque or uninspected alternatives. Known local units do not establish complete system inventories, and no finding asserts observed operation or causal benefit. The retained uncertainty bounds the conclusions faithfully; it is not an additional unresolved verification defect or publication limit.

Blockers

none

Limits

none