Cortex memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The profile classifies accepted records, not a deployment. All positive findings retain implementation wiring or attributed procedure as their basis. No supplied execution establishes exercised maintenance, changed host behavior, improved answers or causal benefit. The partial assessments preserve the memory report's unresolved initial-content, operational-payload and provenance inventory rather than treating one positive route as complete coverage.
Content versus access structures. MEM-OBJ-notes and MEM-OBJ-bubble-body supply interpreted context; MEM-OBJ-bubble-state, MEM-OBJ-note-scores and MEM-OBJ-note-edges supply fixed-rule selection inputs. Postgres edges establish relational storage with graph structure, not a separate graph database. Scalar score arithmetic is symbolic, not parametric. Static RT-OBJ-answer-contract and RT-BAP-maintenance-instructions explain instruction force but are not accumulated memory. Unknown private-host interpretation does not establish an instruction finding.
Admission is mechanism-specific. Software event/graph/cache admission is automatic. Human attention changes and pins have their own manual warrant. Credentialed generic writers, proposal reviewers and batch-script callers do not establish who approved individual content. The groundskeeper's human proposal-review policy remains a claimed manual mechanism alongside claimed automatic filing; it does not narrow the trusted server grant. See RT-RTE-trusted-write, RT-RTE-proposal-admission, MEM-RTE-local-curation and RT-RTE-groundskeeper-maintenance.
Direction follows the operation. RT-RTE-context-delivery returns requested retained content: pull. Its internal selectors also choose parts for downstream context without item-specific requests: push. RT-RTE-reader-answer supplies question-selected notes to the provider, then returns an answer to the requester. Availability/budget filters, project/endpoint identity matches and lexical targeting remain separate signals. Explicit path requests and exact answer-cache lookup are not extra identifier-based pushes. The groundskeeper requests its logs and destination pages; choosing which facts to file is write-side curation, not judgment-based push retrieval. Its bounded date/order rules likewise govern host-requested acquisition, not a separate coarse push. Read-back signal is therefore inapplicable to that local acquisition/filing unit, while the axis remains partial because wired push selectors coexist with unresolved administrative selection.
Learning is not demonstrated improvement. MEM-RTE-usage-selection establishes a wired trace-to-retained-selector-to-later-context chain. Access history has learning authority at its update consumer; derived scores and edges have ranking authority at later selectors. RT-RTE-groundskeeper-maintenance supplies a separate claimed log-to-page/digest chain. Its log consolidation retains raw logs; this is not synthesis merely because a host model performs the selection. Neither chain establishes improved capacity. MEM-CLM-evaluation-gate and EPI-RTE-prediction-evaluation do not install a winning ranker.
Serving rows record tool activity and discrete runtime observations, supporting tool-traces and event-streams for that usage route. They do not establish recorded conversational turns or linked state/action/outcome trajectories. The groundskeeper's dated activity prose has unresolved original content, so its trace source cannot be classified merely from the word log. Cached answers are retained model responses, not automatically experience-fed behavior updates: MEM-OBJ-cached-answer and EPI-RTE-answer-cache license replay, not reader training.
Curation remains local. Note/body revision supports evolve; retained working history leaving open context supports invalidate; bounded age/TTL and score decay support decay; hotter pin overrides support promote. The maintenance procedure separately requests consolidation and historical correction/withdrawal. MEM-RTE-retirement-nomination only nominates: it does not support executed deletion or semantic invalidation. Import collision checks, mirror rebuilds and file relocation do not establish dedup, and generated answers alone do not establish synthesis of new claims in curated memory.
Canonical identity. Reconciliation supersedes EPI-OBJ-note with MEM-OBJ-notes, EPI-OBJ-note-edges with MEM-OBJ-note-edges, and EPI-RTE-retirement-nomination with MEM-RTE-retirement-nomination. Those canonical records support the corresponding content/form/storage, graph/ranking and nomination limits here. The graph-rebuild and usage-selection routes remain distinct, as do current reader replies and retained cached answers. No supersession upgrades evidence strength.
Source-read log. No frozen-source reads were needed or performed. Classification uses only the supplied accepted runtime, memory, epistemic and workflow records and their reconciliation; no records or evidence were added.