Dense-Mem memory profile

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

Comparison rationale

This profile classifies accepted records, not additional source facts. All positive findings are implementation wiring; none claims observed operation or causal benefit. Coverage remains partial wherever retained alternatives, provenance or consumer effects are unresolved. The inventory is broader than answering-agent recall: retry checkpoints, worker context, feedback readers, historical displays and evaluation reloads are later memory consumers too.

Forms, storage and authority

Relational persistence and vector access coexist for MEM-OBJ-search-vectors. Parametric here denotes dense encoding, not model training. Relationship edges and JSON serialization do not by themselves establish graph or file substrates. Text, symbolic controls and vectors have separate consumption paths. Historical statement/rationale display does not classify the opaque payload described by MEM-RTE-administrative-views.

Retained eligibility state has enforcement force; retained text/features can influence ranking under fixed code. Session traces at extraction have conditional learning force independently of downstream knowledge authority. Ontology definitions are interpreted source context, not standing instructions: see MEM-BAP-organization-selection. The static assessment/export prompts and external provider parameters are outside accumulated-memory scope. The afforded skill draft in RT-RTE-skill-export has no established retained installation, so it supplies neither instruction authority nor another trace-learning write.

Write and curation distinctions

Automatic admission and unknown human control coexist. Software/model commits require no human content decision, while generic manager/owner/caller roles do not establish manual control. These are different mechanisms, not contradictory classifications of the same decision. Initial vocabulary and historical admission remain unresolved rather than inheriting the automatic writer's classification: see RT-RTE-ontology-maintenance and RT-RTE-manager-publication.

Session extraction is consolidation even though its semantic faithfulness is unobserved. The epistemic indeterminacy in EPI-OBJ-relationship-proposals does not establish synthesis. Correction evolves an entry; withdrawal invalidates reliance while preserving history; retention erasure/pruning forgets content. Index rebuild, requested model behavior and result deduplication do not establish additional retained curation operations. The fixed MEM-OBJ-conflict-derivation is audit/recovery material, not newly recalled factual support.

Read-back and trace uncertainty

Requested recall remains pull despite automatic ranking. New intake automatically supplies matching retained prior events to the extractor; source-linked vocabulary is automatically supplied to the organizer. These internal push operations have actual identity/lexical targeting and separate coarse budget filters. They do not establish unsolicited host-conversation insertion. Their controls are recorded in MEM-OBJ-session-events, MEM-OBJ-ontology-records and the annotations on RT-RTE-session-intake and RT-RTE-ontology-maintenance.

The wired session path qualifies for trace learning when caller text records experience, because it automatically derives durable material for later consumers. That conditional positive does not authenticate all inputs. The accepted records explicitly leave actual event origin mixed or opaque; no session-log, tool-trace, event-stream or trajectory source can be assigned merely from the session/framework wrapper. Trace-source therefore remains not-determinable, not inapplicable. Feedback snapshots and loaded evaluation traces establish retention/read-back, not an independently demonstrated automatic adaptation route: see MEM-RTE-feedback-recording and RT-RTE-evaluation-suite.

Reconciliation and limits

Reconciliation supersedes EPI-RTE-recall-feedback with MEM-RTE-feedback-recording. All profile feedback citations use that canonical route; EPI-OBJ-feedback-assessment remains a separate content object. No other supersession changes a value.

The retained disagreement about original fragment delivery does not prevent form or direction classification. This profile asserts source-derived natural-language context, not delivery of a complete fragment. EPI-RTE-recall-hydration describes truncation and the ordinary lexical-path support-gating counterexample. Accordingly no full-content guarantee or universal relationship-support gate is used to establish an authority finding. The example's discarded recall/trace responses in WF-CLM-client-delivery establish neither semantic verification nor host answer benefit.

Missing initial control, historical payload schemas, actual trace origin, full diagnostic consumers, complete evaluation import provenance and detailed alternative maintenance/recovery selectors prevent stronger coverage assessments. No bounded absence record is supplied that would justify converting these limits into absent axes or a system-wide trace-learning no. No source files were read for this profile; all classifications use the supplied accepted records and reconciliation only.