Cortex memory profile

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

Comparison rationale

This profile classifies accepted records, not a running installation. All positive implementation findings are wired unless an individual finding is only afforded. No observed operation or causal benefit is supplied. The inventory follows the retained parts in MEM-OBJ-documents, MEM-OBJ-summary, MEM-OBJ-relationships, MEM-OBJ-embeddings, MEM-OBJ-tier, MEM-OBJ-access-count, MEM-OBJ-query-log, MEM-OBJ-retrieval-weights and MEM-OBJ-document-versions. Recovery and importer admission are separately covered by RT-RTE-maintain-knowledge and RT-RTE-direct-import.

SQLite vector BLOBs establish SQLite storage and distributed-parametric form, not a separate vector substrate or retained model weights. Symbolic coefficient JSON is not distributed-parametric learning. Dormant weight/history getters remain in the inventory without invented ranking or historical-recovery consumers. The missing shipped bridge is bounded by MEM-ABS-adaptation-wiring.

Requested context ranking, dossier matching, synthesis selection and alerts are pull operations, even when code chooses the returned subset. Their identifiers, lexical scores and embeddings therefore do not establish push-signal values. This distinction follows RT-RTE-retrieve-context and MEM-RTE-issue-delivery. It says nothing about integrations outside the reviewed service boundary.

Retained counts feed automatic promotion, while the promoted tier is later displayed and preserved by enrichment. The update input has learning authority; its downstream effects are informational and branch-selecting, not a demonstrated hybrid-ranking improvement. This is a wired trace-learning witness under the profile definition, independent of the stronger learning-benefit claim qualified in MEM-CLM-learning-scope. Generic documents may or may not contain traces. Their unresolved original content prevents a system-wide negative or complete trace-source assessment. Query-log retention alone is not another qualifying learning route.

The summary branch requests compression of retained content, so consolidation is afforded; no actual model output establishes faithful reduction or synthesis of a new retained claim. Model-derived relationship meaning remains indeterminate. The transient period narrative is excluded from retained curation, as EPI-RTE-period-synthesis establishes. Explicit revisions are evolution; deletion and additive supersession do not establish a universal history-preserving invalidation policy.

Reconciliation supersedes the epistemic summary with MEM-OBJ-summary, epistemic access adaptation with MEM-RTE-access-promotion, and epistemic state diagnostics with MEM-RTE-issue-delivery. All profile citations use these canonical targets. Distinct epistemic generation, checking and rule-retention records remain available for their separate functions. These identity resolutions do not upgrade fidelity, acceptance, operation or benefit.

No frozen source files were read for this classification. The supplied accepted members and their record passages supply the warrants; no new records, quotations or evidence were added.