EvoOntology memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The inventory is partial, not a file-store-only simplification of the whole system. The accepted memory account and reconciliation retain an undeclared InsightBench analytical-history/checkpoint and fallback-consumption inventory. Each axis keeps that included alternative unresolved rather than inferring its properties from normalized trajectories or ToolCallingChat frames. The relevant canonical records are MEM-RTE-insight-guidance, MEM-RTE-trace-retention and RT-RTE-benchmark-tool-loop. This missing inventory prevents complete classification; it does not defeat the scoped positives.
Reconciliation supersedes the five epistemic semantic-object declarations with MEM-OBJ-term, MEM-OBJ-mapping, MEM-OBJ-relation, MEM-OBJ-constraint and MEM-OBJ-evidence. All profile object citations use those canonical declarations. The valid RT-OBJ-evolution-session container remains alongside its hypothesis, evaluation and source-reference parts; containment does not upgrade their individual warrants.
Semantic JSON links are symbolic access structure, not a graph database. Gate numbers are symbolic observations, not learned parameters. Query/filter/connection strings at the analytical model supply knowledge through interpretation; they are not enforced native-query rules. Retained strings also serve as ranking inputs at fixed lexical selectors. Those independent effects are grounded in RT-RTE-semantic-retrieval, the five semantic families and the three consumer-specific guidance records. No universal withdrawal guarantee follows: MEM-RTE-bird-guidance drops lifecycle state, and MEM-RTE-insight-guidance retains a shortcut/binder inspection limit.
Admission control is classified at the content decision, not at disk I/O. The model-directed Build/Evolve route has automatic admission as a claimed method; deterministic trace/workload derivations supply wired automatic witnesses. Generic supplied-content APIs and initial thresholds remain not-determinable because no explicit per-content human decision is established. Their uncertainty is not erased by another route's automatic writer. See RT-RTE-build-publication, RT-RTE-evolution-admission, RT-RTE-version-override and MEM-OBJ-checkpoint.
Trace-fed workload reuse qualifies because it automatically retains questions/provenance derived from task experience and later shapes Builder/task inputs. The separate model-mediated trace-to-ontology route remains claimed at the update step even though persistence, publication and downstream guidance are wired. Neither classification establishes improved future capacity, successful causal attribution or exercised self-improvement. Raw trace retention alone does not qualify. Tool observations and task-linked experience support the stated source values; discarded agent prose does not establish a session-log learning route. Opaque supplied observations and reference bodies retain their origin limits. See MEM-RTE-workload-reuse, MEM-RTE-trace-retention, MEM-OBJ-hypothesis and RT-RTE-evolution-admission.
Curation distinguishes durable workload deduplication and trace consolidation from reply suppression. InsightBench's transient browse/resolve history is not persistent semantic curation. Semantic revision is prescribed, and direct named-version replacement affords revision, but neither supplies observed semantic novelty. Projection, publication and index selection do not alone establish synthesis, decay or promotion. The opaque transformation unit therefore retains uncertainty rather than declaring those operations absent. See MEM-RTE-workload-reuse, MEM-RTE-trace-retention, MEM-RTE-insight-guidance, RT-RTE-version-override and EPI-RTE-semantic-projection.
Requested task setup remains pull, even when code chooses the supplied manifest. In particular, MEM-RTE-ddr-guidance establishes that shipped callers omit the optional task argument; the optional relevance filter is afforded ranking, not wired task-targeted manifest delivery. The established unsolicited route is MEM-RTE-reminder-delivery. It uses coarse availability/count/time criteria and sends checkpoint-derived reason/action context. A checkpoint task pointer identifies the counting boundary, not an identity match selecting delivered semantic parts. Pull-only units are locally inapplicable to push signals, but do not make the partial system axis inapplicable.
No target source reads were performed for this classification. All warrants come from the supplied accepted records and their annotations, resolved through the supplied reconciliation. No observed run or causal experiment is claimed.