Semantic Engine memory profile

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

Comparison rationale

Canvas label text, geometry/IDs and author-filter routing retain separate afforded findings from the wired archive/SQLite text, symbolic fields and query parent-ID reload. Their controlled values remain supported, but each finding carries its own local route warrant. The qualified canvas producer/consumer in MEM-RTE-canvas-display cannot inherit the stronger basis of RT-RTE-query-documents.

The inventory separates local retained parts from potential Atlas state. SQLite is the substrate even where its payload is a dense vector. Dense vectors receive the parametric form classification without attributing updates to model weights. Actual binary conversion remains unresolved in MEM-OBJ-chunks-embedding; that limits interoperability, not the identification of the ranking consumer.

Reconciliation supersedes the four epistemic document/chunk object declarations with their memory counterparts. This profile cites those canonical memory records. The valid runtime container and the separate canvas/Atlas route parts remain; no supersession upgrades an afforded connected display route to wired operation. The topic request remains an opaque alternative under EPI-RTE-atlas-topics, not evidence of synthesized retained claims.

Explicit vector clearing is manual authorization of removal, while ingestion and embedding admit their results automatically. Clearing forgets the vectors without retaining their history, so decay is narrower and better supported than invalidate. The surviving document text is a different operative part, not the removed vector's history. See RT-RTE-clear-embeddings. Replacing a derived map artifact supports afforded evolution; numerical projection alone supports neither consolidation of claims nor synthesis.

Ranking and root reload occur inside a requested query. They are not push simply because code selects and delivers the response. Likewise, a username-named map file is requested whole, not matched to an automatic context-injection trigger. The known local pull units therefore have no applicable push signal, but opaque remote consumers prevent a whole-boundary inapplicability finding. See RT-RTE-query-documents and MEM-RTE-canvas-display.

Trace learning and trace source remain not determinable rather than importing a system-wide negative from the prose. The accepted set contains no canonical bounded trace-learning absence declaration, and external topic/initialization alternatives are unresolved. The factual-acceptance absence is about a different property and supplies no such warrant. A future positive would need original trace inputs, automatic durable admission and a later behavior-shaping consumer in accepted records. Missing executions independently prevent observed or causal findings; they do not weaken the inspected local wiring into mere claims.

Source-read log: none. Classification used only the supplied accepted members and reconciliation; no target-source read or additional evidence was required.