llm-wiki memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies retained objects and their actual consumers, not every file or model used by the runtime. The supplied records establish filesystem persistence and later-consumer routes, but leave several initial admissions, special-cache routes and included variants unresolved. The positive findings therefore do not claim exhaustive system coverage. No observed or causal basis is assigned: the accepted members supply implementation and doctrine only.
Objects, encoding and authority
Retained graph JSON is a file substrate with symbolic content, not evidence of a graph database. Likewise a floating confidence field is symbolic metadata, not parametric memory. Prose, exact fields and pending UUIDs have different consumers. The rendered pending task gives the foreground agent directions/context; the sentinel gives completion code an enforced association criterion. See MEM-RTE-graph-query, MEM-OBJ-quality-metadata, RT-OBJ-pending-prompts and RT-RTE-agent-completion.
Learning authority is local to the update consumer of retained experience. Knowledge authority is local to the later reader of the resulting page. Human approval of a self-learn diff changes write agency, not the transcript's role at the revision consumer. The trace-learning axis instead requires automatic admission. The manual self-learn unit can consequently carry learning authority without a trace-learning yes. See MEM-RTE-self-learn, RT-RTE-backend-synthesis and MEM-OBJ-wiki.
Transformations and delivery
Summarization/excerpt reduction supports consolidation; it does not establish novel claim synthesis. Opaque model outputs retain their semantic uncertainty. Force replacement and candidate merge support evolution; promotion raises the candidate tier. Archiving is bounded invalidation of trusted-area placement, not corpus-wide suppression: recursive retrieval can still return archived claims without their rejection reasons. See RT-RTE-backend-synthesis, RT-RTE-candidate-admission, MEM-OBJ-candidate-archive and MEM-CLM-archive-visibility.
MCP lexical matching, exact page paths, graph seed matches and foreground relevance judgment occur inside requested read-back. They establish no push signal. Pending prompts are picked up by an explicitly invoked procedure, not a launched worker. The unresolved SessionStart/queue and hot-memory routes prevent an exhaustive pull-only conclusion, so read-back signal is not-determinable rather than globally inapplicable. See RT-RTE-mcp-access, MEM-RTE-graph-query, RT-RTE-agent-completion, MEM-OBJ-ingest-queue and MEM-OBJ-navigation.
The optional reader overview is separately retained HTML derived from session metadata. It is included as a trace-fed retained context route, without claiming that it enters curated wiki Markdown or that a reader actually used it. Source classification follows session origin, not the timestamped serialization or the name of the synthesis operation. See EPI-OBJ-reader-overview and RT-RTE-reader-build.
Canonical identities and unresolved evidence layers
The supplied reconciliation supersedes EPI-OBJ-wiki-pages with MEM-OBJ-wiki, and EPI-OBJ-quality-metadata with MEM-OBJ-quality-metadata. All classification citations use those canonical identities. The supersessions change no values; they retain unknown semantic fidelity and the distinction between transient helper returns and persisted quality fields. The separate reader-overview object and grouped quality/maintenance routes are not superseded.
The supplied reconciliation leaves maintenance implementation status disputed: MEM-RTE-wiki-maintenance says afforded; EPI-RTE-page-update and EPI-RTE-wiki-reflection say claimed. Relevant findings retain both bases instead of upgrading the claimed procedure. Both accounts agree on direct update admission and approval-gated reflection, and neither establishes execution. This limits a unified implementation verdict without preventing the bounded classification above.
Initial candidate/custom navigation content, persisted confidence writers, queue consumers and special-cache selectors remain unresolved. The coverage limits carried by MEM-OBJ-navigation, MEM-OBJ-quality-metadata, MEM-OBJ-ingest-queue and RT-RTE-candidate-admission prevent stronger coverage assessments. Unread repair/vault/editor variants remain included uncertainties rather than evidenced absences. No source files were read for this mapping; only the supplied members and governing contracts were used. No records, quotations or source evidence were added.