Quicky Wiki memory profile

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

Comparison rationale

The authoritative accumulated graph is SQLite, not a vector store or model-weight update. Its text, fixed-rule controls, provenance, relations and reduced event history remain separate units. Files are independently retained surfaces: the standard answer reads the graph, whereas dashboard page display can reload rendered Markdown. Export readers are afforded, not observed. These boundaries follow RT-OBJ-knowledge-store, MEM-OBJ-rendered-views and RT-RTE-question-answer.

Reconciliation supersedes the epistemic statement and summary duplicates with MEM-OBJ-claims and MEM-OBJ-page-summaries. All classifications use those canonical objects. The confidence-state container, SQLite store container and slide-specific parts remain valid overlapping scopes, not replacement identities. No supersession changes a controlled value.

Authority is consumer-relative. Retained text participates in lexical ranking even though fixed code computes retrieval; scores rank review candidates and gate default slide inclusion. That gate is not a universal truth veto: a topic override bypasses it, and ordinary answers have no confidence floor. Hashes, metadata and relations route work. Configured entity guidance supplies instructions rather than ordinary evidence. See MEM-OBJ-claim-controls, MEM-OBJ-source-register, EPI-RTE-slide-output and RT-BAP-retained-context. Host mutation enforcement is not itself evidence that every retained part has enforcement authority.

Requested answers and client reads are pull. Compiler-side automatic provision of existing graph context is a separate push: coarse all-content selection and actual page/kind identity matches coexist. Watch supplies file content to admission, not to an answering conversation. Its path match is a selector of the delivered file, unlike a passive identifier on a requested catalog entry. The record chain is RT-RTE-source-ingest, MEM-OBJ-page-summaries and RT-RTE-runtime-setup.

The answer-save path separates manual retention from automatic re-extraction. The latter consumes recorded question/answer content and writes claims for later answers; it therefore supports wired trace learning and session-log origin without claiming weight training, improvement or observed activation. Saved output is not independent corroboration. Event rows alone do not establish such an update. See RT-RTE-answer-save, RT-RTE-source-ingest and MEM-OBJ-epistemic-events.

Curation findings do not promote requested semantic behavior into verified transformations. Summary compression and semantic duplicate avoidance remain claimed at their model mechanisms. Numeric evolution and decay are wired. Prefix filtering removes transient candidates before admission, not retained memory; index rebuilding is access maintenance. No controlled synthesis value follows from uninspected output meaning. See MEM-OBJ-page-summaries, MEM-OBJ-claim-controls, EPI-RTE-chunk-selection and RT-RTE-source-ingest.

All axes retain provider opacity and route-specific missing coverage instead of inferring absence. Source originals, arbitrary embedding writes, remaining export payloads and provider-side retention limit different units; one inspected graph route cannot resolve them. Missing runs separately prevent operation and benefit claims. See RT-CMP-llm-adapter, RT-RTE-client-access and MEM-OBJ-rendered-views.

Source-read log: none. Classification used only the supplied accepted members and reconciliation; no frozen source reread was needed.