EXIA SEED memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
Inventory and evidence strength
The comparison separates retained prose, fixed-parsed controls and retrieval structures. Files are the durable substrate of the graph; graph topology is its symbolic encoding. The SQLite FTS cache is process-local, not a durable database of independently acquired memories. External model components are consumers/producers, not evidence that EXIA SEED retains changing provider parameters (MEM-CMP-memoire, MEM-OBJ-graphe, RT-CMP-consolidation-model).
All ten axes retain partial inventory coverage. This is not an inference from missing run traces: wired architectural values are classifiable. It follows from unresolved included starting-content/caller provenance and benchmark alternatives where relevant. The known replay modes do not classify opaque baseline embeddings or every fusion transformation. The units preserve these local limits instead of excluding alternatives to obtain complete coverage (MEM-OBJ-thesaurus, MEM-OBJ-registre, RT-RTE-benchmark-replay). No observed or causal finding is inferred from implementation.
Content, controls and update routes
Standing conduct corrections can be instructions at a later host model while ordinary facts and summaries supply knowledge. The same retained experience has learning authority at the automatic update consumer. Retrieval features have ranking authority under fixed code. Sources and vocabularies have validation/enforcement force at admission gates, whereas privacy labels and confidence claims do not thereby become authorization or truth controls (RT-BAP-recalled-content, MEM-OBJ-page-metadata, EPI-RTE-fragment-check, EPI-RTE-summary-check).
Automatic fact extraction and summary compression establish consolidation as a transformation route, not semantic fidelity. Indeterminate output semantics do not establish synthesis, and a possible negation error does not redefine consolidation as synthesis. Summary replacement establishes evolution; normalized duplicate suppression establishes deduplication. Withdrawal invalidates raw-source eligibility and selected graph references, not all downstream knowledge. Page prose survives reference removal (MEM-OBJ-page-facts, EPI-RTE-summary-write, EPI-RTE-withdrawal-read).
Explicit human retention commands differ from software-controlled admission of ordinary turns and model-produced phrases. Note/native caller identity and an arbitrary starting thesaurus do not establish per-content human control. A write-enabled maintenance invocation authorizes a mechanism, not each resulting phrase (RT-RTE-conversation-ingest, RT-RTE-register-admission, RT-RTE-state-maintenance).
Read-back and trace interpretation
Phase-2 prefetch is push because selection is automatic without a foreground memory-tool request. It is query-triggered lexical/graph inference; the subject identifier in the delivered pointer does not alone establish identifier-based push. Explicit search/read and requested page/source/cache loading are pull. Phase 1 returns no automatic content, without disabling explicit tools. Estimated budgets do not prove a universal final-context ceiling (RT-RTE-memory-delivery, MEM-CMP-memoire).
Session traces automatically produce retained pages, summaries and retrieval inputs that affect later context. This supports wired trace-learning "yes" without claiming observed benefit, host compliance, weight training or exercised self-improvement. Original inputs support session-logs; ordinary timestamps, voice substitution or note serialization do not add an event-stream source. Unknown note origins remain separate from that known chain. Audit logs without an update reader are not additional qualifying learning routes; no accepted bounded absence record supports a trace-learning "no" (MEM-OBJ-raw, MEM-OBJ-journals, RT-RTE-page-consolidation, MEM-OBJ-summary, MEM-OBJ-thesaurus, RT-BAP-recalled-content).
Supersessions and source-read log
Reconciliation supersedes EPI-OBJ-summaries with MEM-OBJ-summary, and EPI-OBJ-keywords with MEM-OBJ-keywords. This profile cites those canonical objects. The fragment and concept containers remain valid alongside their differentiated memory parts; they create no extra values simply by overlapping.
No source files were read. Classification used only the supplied accepted members and reconciliation; no evidence or record was added.