AI Research OS Workshop memory profile

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

Comparison rationale

This profile classifies retained research, not all host state. The same YAML file has symbolic access fields and interpreted summary strings; the same chart artifact has executable constants and visual output. Those parts do not inherit one another's form or authority. The reusable clone is a repository substrate, while its curated specifications are files. Clone code/documents supply knowledge to specification writers. The operation log has no established later-model consumer: its proposed chart tally remains unresolved, so behavioral-authority coverage is partial rather than complete. Visual payload form and initial upstream provenance remain unresolved in their own units.

Helper persistence is wired. Model authoring, compression, preservation, semantic checks and later research consumption are afforded by prompts. None is promoted to observed operation or measured benefit. Catalog projection alone establishes neither semantic synthesis nor curation; stale flags alone establish neither invalidation nor decay. The epistemic semantic-change limits prevent treating an argumentative filename as evidence of a newly created claim. See RT-RTE-index-build, MEM-OBJ-synthesis, MEM-RTE-comparison-writing and RT-RTE-lint-maintenance.

A requested answer/export is pull. Automatic selection of retained prior pages for a delegated writer is a distinct push operation inside that requested workflow; it needs no autonomous scheduler. Identity-based prior-page assignment and render input availability retain their push signals. Summary writers request catalog, rolling files and discretionary full pages themselves; judges request qualifying pages. Their availability/budget filtering and discretionary judgment constrain pull and establish no push signal. No push signal is inferred merely from a requested URI. The external appendix's documented writer role does not establish its actual delivery trigger. See RT-RTE-append-ingest, RT-RTE-query-answer and MEM-OBJ-research-appendix.

The reconciliation supersessions affecting classifications resolve source pages, thesis, repository specifications, appendix and comparison admission to their canonical memory records. The overlapping raw/index/topic/render/lint containers are not superseded by individual parts. No superseded epistemic declaration supplies a second existence witness or upgrades a local basis. The unresolved structural-diagnostic/stale-flag record-part issue concerns transient/report-only outputs excluded from this retained-memory inventory; it prevents a complete part-addressed lint account, not classification of the separately declared rolling contradiction/question files. See EPI-OBJ-lint-findings, MEM-OBJ-log, MEM-OBJ-contradictions and MEM-OBJ-open-questions.

Query write-scope conflicts, GitHub copy/cache seams, missing catalog history and same-path render noops retain their source-level limits. They do not erase supported file retention or model admission affordances. Neither a source-backed research update nor a retained operational summary alone establishes qualifying trace learning. Missing recorded trace-input and later-consumer facts prevent a trace-source classification; the trace axes therefore assert neither a positive nor a complete negative. See RT-RTE-query-answer, RT-RTE-github-ingest, MEM-OBJ-log and RT-RTE-render-output.

Source-read log: none. Classification used only the supplied accepted members, their annotations and reconciliation; no source ambiguity required a new frozen-source read.