Antinomia memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies accepted records, not a live vault. Wiring establishes a consumer path, not successful semantic reliance, improved capacity or observed operation. Settings identifiers do not establish immutable model weights. A dependency graph encoded in YAML is a symbolic access structure over files, not evidence of a graph storage service.
The reconciliation supersedes the epistemic inline-assumption route with MEM-RTE-inline-assumptions, and the epistemic seeded-measurement claim with MEM-CLM-evaluation-key. This profile uses those canonical destinations. The evaluation key is not a production trace-fed update consumer. Narrower epistemic objects remain compatible with the broader memory objects; they are not silently substituted for them.
Two unresolved reconciliation issues are kept bounded. The U-prefix deletion condition in MEM-OBJ-note-metadata and EPI-RTE-collapse-review prevents a universal dependency-cleanup assurance; it does not erase the retained link/routing path. Inline A/B content lacks a separately declared operative-object part in the combined epistemic inventory. Its storage, edits and later modal/Hunter consumers are explicitly recorded by MEM-RTE-inline-assumptions, which supports the local classification without inventing that missing declaration or claiming a repaired object inventory.
Human admission and software derivation are separate mechanisms. Human batch selection admits concepts; code separately constructs hubs and backup copies. A human's command initiation is not per-content approval of every derived write. Historical admission of existing vault notes remains unresolved. See RT-RTE-source-ingest, MEM-OBJ-source-hubs, MEM-OBJ-migration-backup and MEM-OBJ-note-content.
Read-back distinguishes operations, not human versus agent consumers. Hunter's automatic type/recency/cap selection fulfills a requested scan; it does not establish a second push merely because code chooses inputs. Initial human view opening likewise requests inspection. Later host-event-triggered refresh automatically reloads retained notes, applies type/status/layer and view filters, and supplies summaries to the human without a new request. That operation supports wired push with coarse selection even though it is not agent-context injection. Settings reload and dismissal-control consumption remain separate unresolved alternatives, keeping both axes partial. Frontmatter dismissal keys select output vetoes, not delivered memory parts, so exact pair matching supplies no identifier push finding. See RT-RTE-hunter-scan, MEM-RTE-retained-inspection, MEM-OBJ-plugin-settings, MEM-OBJ-hunter-feedback and RT-BAP-dismissal-filter.
Curation is not epistemic warrant. Requested faithful extraction supports a claimed consolidation mechanism, while the accepted epistemic account leaves actual concept meaning indeterminate. Generated principles and assumptions do not by themselves establish synthesis. The wired replacement and lifecycle-withdrawal operations remain supported independently. See EPI-CLM-source-fidelity, RT-RTE-rule-admission, MEM-RTE-note-maintenance and EPI-RTE-collapse-review.
No member declares an evidenced absence. Accordingly the trace-learning axis cannot encode a known local negative, and trace source cannot be made inapplicable by a narrative that no qualifying rewrite was established. Raw diagnostics, counters and manually admitted feedback do not establish an automatic trace-fed durable behavior-shaping artifact. This is a limit on classification, not a claim about provider learning or human improvement. See MEM-OBJ-hunter-raw, MEM-OBJ-plugin-settings and RT-RTE-pair-dismissal.
Source-read log: no frozen-source reads were needed. This mapping uses only the handed members and governing contracts; it adds no records or evidence.