expo-llm-wiki memory profile

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

Comparison rationale

Boundary and evidence strength

The profile classifies retained memory and its consumers, not every store used by the runtime. SQLite holds both text and numerical arrays; adjacency and vectors do not independently establish graph-service or vector-service substrates. The local object inventory distinguishes language from symbolic controls and dense numerical representation. Imported summary exposure is not a demonstrated continuation-summary route. See MEM-OBJ-entry-vectors, MEM-OBJ-edges and MEM-OBJ-entity-summary.

Every positive basis is wired except the explicitly afforded prospective task-instruction consumer. No observed-run or causal basis is inferred. ScopeLab supplies inspected request construction and mixed feedback content, but its missing-provider and persistence-property mismatches remain limits on operation. The replay/reopen benchmark is an independent wired trace-to-later-consumer witness, not evidence of successful execution or improvement. See RT-RTE-scopelab-chat and WF-RTE-replay-cache.

Each axis remains partial because the shipped remote integration does not reveal retained payloads or remote admission/selection mechanisms. Host-defined outbox effects and unresolved local initialization or helper branches are kept where they prevent stronger classification. These are not negative findings. Neither external provider model internals nor static configuration is substituted for accumulated memory. See RT-RTE-remote-tools, RT-RTE-outbox-sync and MEM-OBJ-entity-manifests.

Admission, transformation and authority

Explicit human supply of a user message is a manual content decision. Subsequent model extraction, admission and maintenance remain automatic; authoring or spending authorization does not approve their contents. Generic caller-controlled remember/import/reviewer APIs cannot establish human control. Rejected resolver operations can still admit assertions, so operation rejection is not a universal content veto. See MEM-OBJ-events, RT-RTE-direct-admission, EPI-RTE-ops-disposition and WF-RTE-report-delivery.

Event-to-standing-fact extraction supports consolidation without proving semantic faithfulness. Resolver updates, deduplication, temporal withdrawal, lifecycle promotion and age/access decay have distinct mechanisms. Generated newFacts do not establish novel-claim synthesis: the accepted epistemic records leave semantic preservation, entailment and non-entailment unresolved. Retained criticism or improved knowledge is not inferred from heal's name. See RT-RTE-event-librarian, MEM-RTE-temporal-history, EPI-RTE-heal-dispose and MEM-RTE-memory-withdrawal.

Persisted manifest language guides writers while symbolic constraints validate and sometimes veto candidates. Access/trust/validity fields act at ranking, filtering and scheduling consumers; default inclusion of drafts prevents a blanket enforcement interpretation. Task descriptions afford standing action requests, but the benchmark does not deliver them. Trace inputs have learning authority at their durable update consumer; resulting facts have knowledge authority downstream. These forces do not imply improved capacity. See MEM-BAP-manifest-guidance, MEM-OBJ-entry-metadata, EPI-OBJ-task-description and EPI-RTE-grounding-status.

Read-back and trace-source distinctions

A requested read remains pull even when its implementation ranks and packs results. The benchmark's requested memory return is not counted again as push merely because it enters an answer prompt. ScopeLab separately supplies selected context before Gemini asks for memory, and maintenance automatically supplies selected retained context to its providers. These independent selector operations support push. ScopeLab retains a lexical finding only: its accepted records do not establish caller-specific embedding-enabled selection, so the former local wired inferred-embedding finding is withdrawn rather than downgraded to a conditional capability. The separately recorded ops vector selector remains wired. The profile distinguishes coarse windows/eligibility from lexical anchors and vector/keyword neighbors; IDs on requested history or graph objects and resolver aliases do not establish identifier-targeted push. See RT-RTE-memory-read, RT-RTE-benchmark-answer, RT-BAP-memory-context, MEM-RTE-heal-memory and EPI-RTE-ops-classify.

The benchmark original input is session turns, despite observation-event serialization. ScopeLab additionally records tool-call/result traces. Durable access feedback is a separate runtime-event route: read activity becomes counters/times consumed by later ordering and retention. Generic event/document APIs and imported summaries cannot establish original trace content. Supported local yes findings do not resolve those alternatives or license a complete trace-source union; no absence record supports local or system-wide no. See WF-RTE-replay-cache, MEM-OBJ-events, RT-RTE-memory-read and MEM-OBJ-entity-summary.

Reconciliation and source-read log

The supplied reconciliation supersedes the epistemic fact-text declaration with MEM-OBJ-entry-text; this profile uses that canonical object. Task descriptions and grounding trust remain valid parts of the memory containers, and runtime route containers coexist with their specialist functions. No other supersession changes a value. The ops and benchmark refinements resolve previously open details rather than upgrading evidence strength.

No frozen source file was read for this classification. All warrants are accepted member records and their supplied annotations, resolved through the supplied reconciliation. No records, quotations or new source evidence were added.