llmwiki memory profile

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

Comparison rationale

The inventory separates retained knowledge from access structures and from the static policy that writes it. MEM-OBJ-wiki-pages is the changed-through-use product; the shipped save skill is not counted as an accumulated instruction store. Provider model components are not evidence that this memory system writes parameters. Session-local term and seen-state still count where later invocations consume them, while current-call grounding cards alone are not an additional durable store.

The accepted reconciliation supersedes the epistemic page, quiz and topic-merge duplicates with MEM-OBJ-wiki-pages, MEM-OBJ-quiz-ledger and MEM-RTE-topic-merge. These canonical declarations are used for those referents. The narrower EPI-OBJ-edit-rule overlaps, but does not replace, the broader rule/recall extension. Feedback and review containers are not flattened into a pooled consumer or stronger evidence basis.

The same rule string has two consumption paths. MEM-OBJ-recall-triggers is symbolic lexical input to ranking, even when imperative delivery is withheld. MEM-OBJ-wiki-rules separately supplies natural-language instructions and path targeting. The RT-RTE-edit-rules can produce a scoped first-edit denial, not proof that a host followed a rule. This distinction supports ranking, instruction and enforcement without equating them.

An operator starting save or enabling commit does not approve every admitted claim. RT-RTE-warm-save, RT-RTE-auto-update and MEM-RTE-topic-merge establish routine model/software admission. Human direction answers and MEM-RTE-query-fileback are separate manual policies. Generic verification APIs do not identify a human decision merely by accepting a caller string.

The profile classifies reduction of retained session content as consolidation and topic replacement as evolution. It does not infer semantic synthesis from the word “synthesis”: MEM-RTE-topic-merge and the MEM-RTE-query-fileback do not establish non-entailment of an actual new output claim. The distill check's narrow source/conflict preservation and the merge route's failed-replacement recovery limit remain limits on fidelity, not reasons to suppress supported curation operations. MEM-RTE-tier-cleanup changes availability while preserving bodies; feedback withholding is version-bound.

RT-RTE-context-delivery and its memory annotation support automatic cold and lexical turn supply. Requested page reads are pull, not a second push inside that delivery. RT-RTE-edit-rules supplies an actual target-path match for identifier-based push. Pointer-only delivery is not a demonstrated later page Read, and bounded excerpts omit evidence lines and footnote definitions. The scope therefore does not assume every delivered claim carries retained reasons.

Automatic trace-fed production of durable wiki context qualifies for trace learning even if the content only summarizes prior experience. MEM-RTE-feedback-fold independently qualifies through durable selector adaptation. The raw-input learning authority and downstream knowledge/ranking authority are separate consumer paths. Wired trace learning does not assert observed improvement, model-weight updates or exercised self-improvement. Legacy read-scan refresh is explicit, not a default running daemon learner. The profile asserts no local negative because the accepted reports declare no bounded evidenced-absence records establishing one.

Every axis preserves unresolved coverage rather than excluding an inconvenient branch. The accepted RT-RTE-context-delivery names commit and absence channels, but records do not inventory their retained parts or evaluator domains. Initial page/recall provenance, generic caller control, some maintenance transformations and observer inputs impose additional local limits. These gaps prevent a complete memory-value inventory, not the positive wiki, rule, feedback and quiz classifications. No full-history absence claim follows from an absence-channel label.

No source files were read for this profile. All classification warrants come from the supplied accepted records and their annotations; no new evidence or records were introduced.