Semiont memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This revision-2 comparison classifies the accepted records, not a running deployment. Every axis remains partial: inspected paths establish positive findings, while included host, admission, content and backend alternatives remain unresolved. Wired means an inspected mechanism, not observed activation, factual accuracy or causal benefit.
The resource file, historical event envelope, annotation prose, symbolic anchors, graph structure, vector dimensions and readable search payload are different operative parts. A KB working-tree file is classified as files, not repo merely because Archivist stages it. Queue-backed retained jobs are service objects; no unrecorded storage engine is inferred. External inference weights are not accumulated KB memory. See MEM-OBJ-content, MEM-RTE-record-replay, RT-OBJ-job-params and MEM-OBJ-vector-points.
Lineage describes acquisition or derivation, not truth or human approval. Direct bytes are imported, model compositions are authored, and deterministic replay/extraction/embedding is other-compiled. Outcome/progress controls are trace-extracted at their own update paths. Generic caller identity cannot establish manual admission; the inspected Worker and follower writes are automatic. The uninspected restoration alternative cannot inherit live-command guarantees. See RT-RTE-resource-admission, MEM-RTE-content-revision, RT-RTE-mark-detection and MEM-RTE-vector-maintenance.
Knowledge supply, candidate ranking and maintenance routing are independent forces. Retained assignments/vocabulary can also validate and enforce admission. Assessment prose is advisory knowledge, not standing instruction or an automatic truth veto. Learning authority of an extraction outcome at its updater differs from routing authority of the resulting remembered decline at startup. These mappings do not assert learned model weights or improved capacity. See MEM-OBJ-annotation-text, MEM-RTE-reference-search, MEM-BAP-projection-control and RT-RTE-resource-admission.
The reconciliation supersedes the epistemic reference-search declaration with MEM-RTE-reference-search. All ranking findings use that canonical route; supersession does not strengthen evidence. The retained disagreement over the workflow's unqualified citation/completion sentence is not used as evidence of successful citation storage. RT-RTE-yield-generation and EPI-RTE-citation-retention establish refusal of this caller's nonempty generated-citation batches. Admitted resource bytes and successful source binding/linking remain separate retained outputs. No profile finding counts those refused citations as memory.
Requested retrieval remains pull even when the service uses embeddings or a model ranker and the Worker clips the response before prompting. Automatic event-targeted maintenance supplies selected retained parts independently of a knowledge-retrieval request; its identifier and coarse progress/compatibility selectors are classified separately. This is not evidence of unsolicited foreground-agent prompt injection. See RT-RTE-annotation-gather, MEM-RTE-resource-gather, MEM-RTE-graph-maintenance and MEM-RTE-vector-maintenance.
Content/body revision supports evolve; withdrawal with retained event history supports invalidate; terminal queue pruning supports local decay. Replay and reindexing do not themselves establish semantic consolidation, deduplication or synthesis. Summary/generation requests leave actual meaning changes unresolved, and aggregate composition remains afforded. See MEM-RTE-annotation-edit, RT-RTE-mark-detection, EPI-RTE-summary-generation and MEM-RTE-aggregate-writing.
The trace-learning positives are narrow: extraction outcomes become durable decline controls, and applied/completed activity becomes durable resume controls read at later startups or attempts. These are behavior-shaping artifacts even without new factual claims. Their source is runtime event/progress content, not sessions merely because an agent participates. Other generation and projection routes retain unresolved trace qualification; imported document content does not become an event-stream source simply by entering the event log. See MEM-OBJ-declined, MEM-RTE-vector-maintenance, MEM-OBJ-weaver-checkpoint and RT-RTE-mark-detection.
Source-read log: none. This profile adds no source evidence or declarations and does not reopen report verification. Missing facts stay beside their scoped findings rather than being filled from unrecorded source facts.