kgai memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The inventory separates retained reasons from the symbolic machinery selecting their current authority. MEM-OBJ-decision supplies knowledge at later model/human readers; MEM-OBJ-event and MEM-OBJ-graph supply structural state and routing inputs. Text can also influence ranking without being delivered. MEM-RTE-context-recall specifically retains this distinction for historical text versus current head reasons. The static instruction to respect or supersede reasons is not itself accumulated memory.
Reconciliation supersedes the epistemic decision-reasons declaration with MEM-OBJ-decision, and the epistemic historical-replay declaration with MEM-RTE-asof-recall. Those are the canonical citations here. The narrower authority part EPI-OBJ-decision-mutations remains distinct from the encompassing event. EPI-RTE-recall-shaping remains a valid overlapping operation; its semantic-preservation uncertainty does not remove wired pull delivery or ranking. No supersession upgrades a basis or changes an axis value.
Read-back direction follows each operation, not the presence of a selector. Lexical context/search and editor requests are pull. SessionStart convention injection is push: configuration precedence and approval choose retained directions without a task request. The approval fingerprint is an actual matched identifier selecting an eligible project layer, not an identifier attached only to a requested file. Coarse layer availability remains a separate selection basis. Neither synchronization nor a static Stop reminder establishes automatic relevance-based decision-history delivery (RT-RTE-config-admission, MEM-OBJ-prompt, RT-RTE-team-sync).
Curation findings concern the implemented change. Successor authority and edge retirement invalidate current reliance while preserving history; live mutations evolve existing entries. Exact content-identity admission no-ops and event-hash deduplication do not establish semantic deduplication. Rebuild is another materialization, not consolidation. Model consolidation retains the weaker policy/affordance warrant because exact inputs and meaning preservation are unresolved; no synthesis is inferred merely from a model-generated rationale (MEM-OBJ-event, MEM-OBJ-graph, RT-RTE-fork-capture).
Trace learning here means an automatic trace-fed write with a durable behavior-shaping result and later consumer, not demonstrated improvement. The fork's parsed decision crosses into persistent event memory and later recall, whereas a pending or skipped proposal does not. Its supported human-confirmation mode is separate from auto and headless fallback. The exact source content copied by the host remains unknown. Eligibility based on successful editing cannot establish that recorded tool traces were consumed, and derived event serialization does not establish event-stream origin. Thus positive trace-learning routes coexist with a not-determinable trace-source axis (RT-RTE-fork-capture, RT-RTE-foreground-capture).
The external model's recorded parametric form is not a warrant for its storage, provenance, admission or learning routes. Those unresolved alternatives prevent complete-value claims without negating known kgai file/graph routes. Generic shell control likewise cannot establish manual admission. All architectural positives remain source-grounded rather than observed or causally supported; no deployed compliance, capture success or benefit is claimed (RT-CMP-host-model, RT-RTE-foreground-capture).
Source-read log: none. Classification used only the supplied accepted members and reconciliation; no frozen-source ambiguity was resolved by additional source reads.