Instinctual Memory memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies accepted records rather than adding source findings. No target source files were read for this job. All positive bases are wiring or explicitly afforded consumption; no observed operation or causal benefit is inferred.
The memory boundary is the target's accumulated material, not its entire storage surface. Cached cross-encoder weights, configurable remote models, installed static skill and setup configuration are inference/deployment machinery, not memory accumulated through the reviewed use loop. Transient proposals and score reports do not become retained memory merely because a later fact can be published.
The published fact and its source reference have separate consumers. Statements reach extraction and host context; evidence quotes ordinarily do not accompany delivery. Suppression reasons and dispositions are readable retained text without an established automatic criterion-revision reader. Their encoding is known while their additional behavioral authority is not determinable. Knowledge, instruction, learning, ranking, routing, validation and enforcement therefore remain consumer-specific rather than a property assigned to every stored fact. See MEM-OBJ-fact, MEM-OBJ-source-ref, MEM-OBJ-controls and MEM-OBJ-dispositions.
Automatic software admission is distinct from a human supplying or approving that content. Command initiation, caller labels and tidy keep exclusions do not settle every human contribution. Tidy's recomputed application is not an approved immutable preview. These limits retain known automatic writes without asserting manual agency from an intended host interaction. See RT-RTE-immediate-change, RT-RTE-tidy-maintenance and RT-RTE-task-update.
Standing-fact extraction is consolidation even where a regex drops negation or a model's semantic fidelity is unresolved. Neither defect establishes synthesis. Corrections revise entries; retirement/soft suppression withdraw reliance with history; destructive erasure forgets targeted content through sequential stages. No complete erasure result is inferred. See EPI-RTE-rules-extraction, EPI-RTE-llm-proposal and RT-RTE-erasure-maintenance.
Push classification separates requested returns from unsolicited selection. Hooks independently select preferences or prompt-matched facts. The extraction chain loads state and automatically supplies a bounded subset to its extractor. Writeback persists a requested generated block, but subsequent host loading remains an affordance with an unresolved selector; the requested target filename alone is not evidence of identity-based push. See RT-RTE-host-hooks, RT-RTE-journal-consolidation and RT-RTE-instruction-writeback.
Trace learning means automatic trace-fed production of durable behavior-shaping material, not parameter training, truth acceptance or measured improvement. Session messages and calendar/voice experience qualify on their recorded paths. Project instruction imports do not become traces merely because they are wrapped as events. Raw IDE/tool retention and retrieval do not establish a qualifying tool-trace update, and generic file/caller origins prevent complete trace-source coverage. See MEM-OBJ-journal-event and RT-RTE-journal-consolidation.
Reconciliation supersedes EPI-OBJ-fact with MEM-OBJ-fact, EPI-OBJ-source-ref with MEM-OBJ-source-ref, and EPI-RTE-index-rebuild with MEM-RTE-index-rebuild. The profile cites only those canonical declarations. The amendments preserve unknown source warrant and bounded metadata derivation; they change citation identity, not evidence strength or value. The valid consolidation container remains alongside its rules, model and candidate-gate parts. No current unresolved inter-report conflict requires selecting a stronger finding.