Voiden memory profile

Type: agentic-system-analyses/types/agentic-system-memory-profile.md

Comparison rationale

The comparison covers memory produced or changed through use, not all storage in the runtime. The inventories distinguish a response record, a binding snapshot, raw outcome history and provider-defined verification state. Their later consumers differ. Installed guidance is included only as the use-refreshed collection and its access index; unchanged shipped prose and downloaded executable capabilities are not counted as learned memory. The record-backed inventory is in MEM-OBJ-response, MEM-OBJ-process-env, MEM-OBJ-session-results, MEM-OBJ-verify-entry-cache, MEM-OBJ-tool-status-record and RT-BAP-agent-guidance.

There are no supersessions affecting these values. Reconciliation retains valid containers alongside their separately assessed parts: runtime variables versus durable JSON, and served decisions versus cache. The CLI status writer and verification consumer overlap without replacing one another. This profile cites each retained canonical declaration rather than assigning the broader container's warrant to its opaque parts.

The strongest positive trace-learning witness is automatic request/response extraction, durable standalone snapshot persistence and later named-value substitution: MEM-RTE-runtime-capture, MEM-OBJ-process-env and RT-RTE-session-driver. That is wired trace-fed behavior shaping, not demonstrated improvement, reflection or a methodological theory update. Its original input is tool request/result material, not session logs merely because the CLI uses session mode. Raw history reporting has a concrete later consumer but no qualifying behavior-shaping derivative in its inspected route: MEM-RTE-history-report.

Selected retained bindings provide symbolic request inputs. Response blocks provide callback context and afford file inspection, but unavailable plugin semantics prevent attributing enforcement or instruction authority to their contents: MEM-RTE-response-reload. The core decision-state comparison has routing force; provider registration, verification/cache use and status read-back are separate unresolved effects in RT-RTE-tool-publication. A readable description note does not classify the opaque consumed runResult. No retained validation or learning authority is inferred merely from status labels or the positive trace-learning route.

Curation includes replacement of a matching retained response. Initial selection/truncation reshapes caller output but does not independently establish consolidation of prior retained memory. These distinct transformations are documented in EPI-RTE-response-serialization and RT-RTE-result-write. Explicit deletion/expiry counts as forgetting, not rollback of remote effects. Deterministic guide summaries consolidate existing retained guide content, not a synthesized claim. Provider status forwarding reduces current output to the upsert payload, but actual retained-memory curation and storage remain unresolved in MEM-RTE-status-write.

Requested file inspection and history reporting are pulls. Supplying selected-section historical blocks to callbacks is a distinct internal push with coarse scope, not identifier targeting merely because the user requested a named file. Binding substitution is identifier-targeted push; whole-map reload is coarse. Provider dynamic-mode search, cache entry selection and host guide injection are unresolved, so local pull-only operations cannot make the signal axis inapplicable. See MEM-RTE-response-reload, MEM-RTE-runtime-capture, MEM-RTE-history-report and RT-OBJ-serve-decisions.

Every axis remains partial because its included inventory contains unresolved alternatives. Evidence strength remains attached to the local mechanism: implemented transformations and delivery are wired; documented file inspection, guide loading and warning interpretation are afforded. Neither means observed activation. Generic caller identity prevents classifying response content as manual or automatic; starting a workflow does not turn software-generated bindings/history into manual writes. Missing provider implementation, actual initial/project content, plugin script semantics and external host context policy prevent stronger coverage. Those gaps are retained rather than treated as bounded absences.

Source-read log: no frozen source files were read for this profile. Classification uses only supplied accepted records and their annotations; no new evidence, quotes or records were added.