Fractal memory profile

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

Comparison rationale

The operative memory is not the whole workspace or runtime storage surface. Session summary, detailed history and usage are contained parts of the checkpoint with different consumers. Changed root guidance, selected configuration and startup cache have their own later-consumer routes. The public API's unspecified supplied context remains an unresolved alternative rather than being assigned the shipped driver's retention protocol. External provider and sandbox internals are outside this comparison, not evidenced absences.

The summary combines interpreted prose with symbolic status/count fields. Its model-facing force is historical context, not standing instruction merely because the signature contains it. Root guidance does supply standing directions. Configuration routes adapter/default choices. Numeric usage informs host display. None of these channels establishes compliance or semantic task acceptance. See MEM-OBJ-session-summary, MEM-BAP-summary-context, RT-BAP-workspace-guidance, RT-RTE-config-change and MEM-OBJ-turn-usage.

The narrow trace-learning yes is automatic production of durable status/accounting/context followed by later model delivery. It does not mean parameter learning, semantic lesson synthesis, accepted truth or improved capacity. Raw trace storage and standalone usage display do not independently qualify. Conversational turns, discrete runtime hook callbacks and execution observations support the positive source findings; complete external RunTrace content and provenance remain unresolved. See MEM-RTE-checkpoint-update, EPI-RTE-event-projection, MEM-OBJ-session-history and MEM-RTE-session-context.

Curation is not assigned from the word summary. The recorded renderer copies entire responses; activity-count projection consolidates detailed turn activity into a reduced representation, even while raw detail remains available in the recent tail. This does not establish semantic abstraction of response prose or novel lesson synthesis. Existing turn finalization and configuration/cache replacement support evolution; recent-tail eviction and cache fallback expiry support decay in their respective scopes. See MEM-OBJ-session-summary, MEM-RTE-checkpoint-update, MEM-RTE-session-context, RT-RTE-config-change and EPI-RTE-version-notice.

Explicit resume is pull. Later submit independently pushes selected-session context. The requested session ID does not make that subsequent all-turn/recency push identifier-targeted; a fixed root filename is likewise not an identity-matched selection of delivered memory parts. Foreground exact-detail inspection is only afforded. The audit/repair recipe demonstrates the shipped continuation wiring, not exercised recall or successful repair. See MEM-RTE-session-reload, MEM-RTE-session-context, RT-BAP-workspace-guidance and WF-RTE-audit-repair.

Reconciliation supersedes EPI-OBJ-session-summary with MEM-OBJ-session-summary. This profile uses the canonical memory object throughout. No other identity is superseded: checkpoint, answer fields, usage part, rendering route and context-delivery route retain their separate scopes. No value is strengthened by that identity amendment.

All positive findings are implementation wiring or explicitly marked affordances. No observed operation or causal experiment is supplied. Partial assessments retain the unresolved direct-API admission/provenance and opaque trace parts rather than projecting a strongest witness into complete system coverage. There are no system-wide absence findings or local trace-learning no findings. Source-read log: no source files were read; classification used only the supplied accepted members and reconciliation.