LLM Wiki memory profile

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

Comparison rationale

The comparison covers accumulated client-side artifacts and their consumers, not every store used by the runtime. The retained file/SQLite inventory is supported by MEM-OBJ-normalized, MEM-OBJ-extracted, MEM-OBJ-wiki-body, MEM-OBJ-wiki-metadata, MEM-OBJ-newsletters, MEM-OBJ-ingested-ids, MEM-OBJ-eval-set and RT-OBJ-metrics. The model services are external dependencies, not evidence of accumulated parameter memory. Static prompt instructions and the process-local registry are excluded. Historical access is retained as a conditional alternative: RT-RTE-api-mutation does not turn article saves into Git checkpoints.

Reconciliation supersedes EPI-OBJ-eval-set with MEM-OBJ-eval-set; all fixture findings cite the latter. No other supersession changes a value. The normalized body and full normalized document, extracted claim/evidence parts and full sidecar, and broad wiki-body/frontmatter accounts remain valid overlapping parts and containers. They are not independent stores or independent corroborating evidence.

Natural-language and symbolic findings describe consumption, not extensions. The same wiki text supplies interpreted model context through RT-BAP-wiki-context, BM25 ranking input through RT-RTE-wiki-consumption, and literal backlink matching through RT-RTE-api-mutation. Confidence, status and source count route fixed-rule maintenance, but do not establish factual warrant or quarantine. Monitor ID rows and sidecar existence do enforce bounded admission decisions. Their access checks do not establish memory-content deduplication.

The browser editor establishes manual content replacement independently of the generic API-client branch. Software admission remains automatic for extraction, integration and maintenance even after an operator launches a cycle. Initial curated-table entries, fixture history, external commits and generic caller identity remain unresolved under MEM-RTE-curated-poll, MEM-OBJ-eval-set and RT-RTE-api-mutation.

Curation positives are restricted to implemented reductions, revisions and promotions. EPI-RTE-claim-projection and EPI-RTE-newsletter-context reduce retained inputs without adding claims at those edges. Subsequent model semantics remain unresolved; a request for novelty does not establish synthesis. EPI-RTE-confidence-update raises a reliance label without new corroboration, while EPI-RTE-newsletter-admission promotes derived output to verified/high-confidence standing after a length gate. Conflict annotations do not establish withdrawal, and MEM-RTE-rss-reset resets admission history rather than dependent knowledge.

Read-back separates explicit material requests from automatic downstream supply. Search returns remain pull even though BM25 selects results. A merge receives an automatically selected old destination body; newsletter, paper, maintenance and optimizer consumers receive code-selected retained inputs. Topic matching is lexical, recentness/category/budget selection is coarse, and destination equality is identifier selection. Upstream operator or model topic choice does not replace the actual retained-page selector in RT-RTE-wiki-consumption.

The trace-learning positive is deliberately local. RT-OBJ-metrics and RT-RTE-prompt-optimization establish automatic execution-feedback processing, retained experiment/hypothesis/decision state and later cycle dependence. Those artifacts shape optimizer behavior even though the proposed prompt is not installed. They support learning authority at that update consumer and a wired trace-learning route, not parameter training, improved future capacity, successful self-improvement or a candidate-to-production-method loop. Incomplete metric producer provenance prevents complete coverage. Recorded fixture subprocess/JSON feedback supports tool-traces on its own branch; no session-logs, event-streams or trajectories value is inferred from timestamps or static source documents. Ordinary page feedback is not silently reclassified as runtime experience.

No new records, source quotations or source evidence were added. No frozen-source files were read for this classification; the supplied accepted records and their annotations supply the warrants. All positive implementation findings remain wiring or affordance, not observed operation or causal benefit. The profile's local coverage assessments do not upgrade those evidence bases.