Beever Atlas memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies the accepted records only. All positive findings are implementation wiring, not observed activation or causal benefit. Each axis is partial because the retained inventory explicitly includes incompletely traced alternatives. Known local units do not close those alternatives.
The reconciliation supersedes only the duplicate contradiction-model component with MEM-CMP-contradiction-detector. Neither provider-owned component is an accumulated-memory object in this scope, so that amendment changes no positive value. Valid containers and their parts remain separately available: the extraction route, callback admission and persisted fact/graph parts are not replacement identities.
Dense retained vectors are parametric representations, not model weights. MongoDB document collections are service objects, not an inferred relational store. The retained registry and alternate adapters remain included without inventing extra substrates from the overall runtime's dependency list. See MEM-OBJ-text-vector and MEM-OBJ-graph-entity.
Push and pull are operations, not whole-system labels. Tool requests remain pull even when an embedding or lexical reranker selects returned facts. Host history injection, event-maintenance context and retry restoration are separate automatic supply operations. Their identity matches and coarse eligibility or budget filters are retained separately. See MEM-RTE-knowledge-retrieval, RT-RTE-dashboard-ask, RT-RTE-wiki-maintenance and RT-OBJ-session-state, including the supplied memory/workflow annotations on those runtime referents.
Retained experience can have learning authority at an update consumer while the derived artifact has knowledge authority at later answering. This mapping asserts the wired trace-fed artifact route, not improved future capacity, training, reflection or self-improvement. Raw log retention alone is not the positive witness: the qualifying QA path includes generated retained answers and later consumption. Original platform/import origins and mixed tool inputs remain unresolved rather than being labelled event streams from their adapter. See EPI-OBJ-qa-answer, MEM-OBJ-chat-history and RT-RTE-message-ingestion.
Consolidation describes the implemented extraction/reduction and summary routes; uncertain semantic fidelity does not establish synthesis. The narration qualifier-removal example does not turn a faithfulness defect into a synthesize finding. Full compiler semantics remain unresolved. Decay denotes forgetting by covered hard purge, not a claimed age-based downweighting rule. See MEM-RTE-summary-retention, EPI-RTE-narration-rewrite, MEM-RTE-wiki-refresh and MEM-RTE-channel-withdrawal.
Coverage limits matter independently of evidence strength. Caller-selected imports, curation and model changes do not prove human admission. Compatibility wiki overlay is not proof that rewritten sections reach every consumer. Fact invalidation excludes on guarded branches but not ordinary BM25 fallback. Diagnostics do not establish an untraced rating-to-policy loop. See MEM-CLM-import-admission, RT-RTE-page-curation, MEM-OBJ-wiki-page, EPI-RTE-fact-supersession and MEM-CLM-evaluation-authority.
Source-read log
No frozen-source files were read for this profile. Classification used the supplied accepted reports and reconciliation; no new records, quotations or evidence were added.