EchoWiki memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The comparison covers accumulated content and its access state, not every storage or permission surface in the runtime. Stored operator edits belong to this boundary because later compilation reloads them. Static fallback instructions, current-call plans, task arrays and shipped model components do not become accumulated memory merely because they affect generation. Opaque provider retention remains an unresolved alternative; the records do not establish that its numerical computation state is a use-generated memory object. The local inventory is supported by MEM-OBJ-raw, the page and metadata objects, MEM-OBJ-index, MEM-OBJ-log, MEM-OBJ-sources, RT-OBJ-agents and RT-OBJ-pending-queue.
Reconciliation supersedes EPI-RTE-page-consume with MEM-RTE-retained-selection. All comparison findings for that progression therefore cite the latter. The page objects and metadata remain distinct valid parts alongside the broader epistemic objects; no other supersession affects a value. No record is newly declared here.
Read-back is split at request, selection and delivery boundaries. The memory report calls the chain request-driven because a compilation starts it. That does not mean the planning or generation model requested retained context. Compiler code automatically enumerates briefs and supplies selected old bodies to those consumers. This supports push with coarse eligibility, planner judgment and target-path matching at their respective operations. The path match selects an actually delivered old body, unlike an identifier merely listed in a catalog. Conversely, preview/current-file reads and documented browsing fulfill requests and remain pull. These distinctions derive from MEM-RTE-retained-selection, RT-RTE-document-compile, RT-RTE-watcher-approval and MEM-OBJ-index, not an assumed query agent.
The planner's retained inputs supply knowledge and help route task choices. No accepted record establishes a relative-priority scoring mechanism, so no ranking value is added. Link validation and enforcement are narrower: retained names help form a whitelist and cleanup changes link syntax. Planned targets enter that whitelist too, so it is not a target-existence or truth guarantee (EPI-RTE-link-check, EPI-RTE-link-reshape). Stored instructions have instruction authority, not an enforced manual-edit veto (RT-BAP-compiler-instructions).
Preview compression establishes consolidate on an inspected branch. Page and catalog replacement establish evolve, without asserting semantic preservation. The requested cross-document synthesis does not establish a claim absent from inputs; the accepted epistemic generation records retain an indeterminate semantic relation. Consequently synthesize is unresolved rather than inferred from the product's terminology (MEM-RTE-page-maintenance, EPI-RTE-page-generate, EPI-RTE-summary-rewrite). Source pointers do not establish a populated archive or evidence recovery (MEM-OBJ-sources, MEM-OBJ-page-metadata).
Automatic page admission and later use are established, but the original source text is not established as traces. Generic voice transcripts do not imply agent session logs or event-stream cues. This prevents either a positive trace-learning route or a complete negative. Log-to-guidance absence stays local, while event-derived pending entries route work without learning from it. Trace lineage for these access/status entries does not imply trace learning (MEM-ABS-feedback-consumer, RT-OBJ-pending-queue, MEM-OBJ-raw).
The reconciliation's failure-retention conflict does not prevent the queue's symbolic form, automatic insertion, routing and reload classifications. It does prevent a universal retry-retention or queue-removal-as-delivery conclusion. Neither is asserted here: the workflow annotation limits the sequential account when another approval returns early on an in-flight path. No classification depends on selecting one report's stronger guarantee.
All positive evidence is wiring or documented affordance. No observed run or causal result is supplied. Initial page/source provenance, unspecified helper callers, opaque media consumption and provider state prevent stronger coverage assessments. This profile does not reopen verification or infer quality, activated learning or improved capacity from persistence.
Source-read log: none. Classification used only the supplied members and their accepted records; no target source file was opened.