o-o memory profile

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

Comparison rationale

This profile classifies the accepted record inventory, not an observed update. Local persistence and shell/browser consumers have wired support. Agent research, source-validity selection and later cache use remain requested behavior; wiring the CLI request does not upgrade those paths. See RT-RTE-document-update, EPI-RTE-web-research and MEM-BAP-cache-reuse.

The reconciliation supersedes the epistemic article, manifest, changelog and contract objects with their memory declarations. Those canonical destinations are used here. The cache container remains distinct from its source and fact parts; the image route spanning staging and embedding remains distinct from persistent image payloads. See MEM-OBJ-article, MEM-OBJ-manifest, MEM-OBJ-changelog, MEM-OBJ-contract, EPI-OBJ-source-cache and MEM-OBJ-images. No supersession changes an evidence basis or establishes actual production provenance.

The substrate union is files plus browser key/value storage, not repo merely because the evidence was captured in Git. Image encoding does not establish its consumed representational form. Readable captions are classified without claiming inspection of the image payload. See MEM-OBJ-images and MEM-OBJ-oo-theme.

Manual supply and automatic admission coexist within creation: the operator chooses topic/settings while code emits default state. Research admission is automatic even when the operator starts the workflow. Operator contract changes and selected machinery replacements remain manual. See RT-RTE-batch-create, RT-RTE-contract-edit and RT-RTE-code-sync.

Index excerpts consolidate retained material, even when truncation loses a qualification. Such a faithfulness defect does not establish synthesis. Research instructions to synthesize an article do not establish the controlled synthesize operation, which requires a claim absent from the inputs. No linked inputs/outputs settle that distinction here. See EPI-RTE-index-excerpt and EPI-RTE-web-research.

Pull delivery and push selection are separate operations: requested document or index viewing is pull, while code selects prominent cards and automatically replays saved presentation state. An eight-card/date filter is coarse; an actual localStorage key lookup is identifier selection. A filename in the agent's read request is not another push signal. See RT-RTE-library-read, MEM-OBJ-index-content, MEM-OBJ-oo-theme and MEM-BAP-cache-reuse.

Trace learning is not inferred from research reuse or the label changelog. The retained summaries lack authenticated execution provenance and an established trace-fed update chain. No accepted bounded-absence record supports a controlled no either. Host compaction could be material but is uninspected; a known local file route cannot settle it. The resulting trace-source uncertainty is not inapplicability. See MEM-OBJ-changelog, MEM-OBJ-facts and RT-CMP-claude.

Source-read log: none. This classification used only the supplied members and reconciliation; it introduced no source evidence or records.