Dynamic Cheatsheet memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The comparison concerns retained material and its access paths, not everything stored or executed by the runtime. The sheet, raw solution archive and extracted round answers are distinct retained parts with different consumers and horizons. Their envelope is classified only where fixed fields carry recovery/access effects. Ordinary adapter chat accumulation and its reset/set affordance do not establish a shipped retained-memory reload route (MEM-OBJ-cheatsheet, MEM-OBJ-final-output, MEM-OBJ-previous-answers, RT-OBJ-output-dict, MEM-RTE-history-controls).
Reconciliation supersedes the epistemic sheet and generator-output object declarations with MEM-OBJ-cheatsheet and MEM-OBJ-final-output. All classifications above use those canonical objects. The epistemic admission and curator-review routes remain distinct accepted mechanisms, not superseded duplicates. No other supersession affects a value.
The sheet's downstream authority is reference knowledge, while experience has learning authority at the curator that revises durable guidance. Neither establishes reliable compliance or improved capacity. Numerical vectors supply ranking; the saved checkpoint and container identifier supply routing. The static tag parser's enforcing replacement rule is not a retained validation criterion, and target-based scoring does not feed a memory veto (RT-BAP-reference-context, RT-RTE-cumulative-update, RT-RTE-context-selection, EPI-RTE-sheet-admission, EPI-ABS-score-feedback).
Whole-sheet replacement supports wired evolution. Consolidation and deduplication remain requested semantic work. A generalization request is not evidence that an emitted claim is absent from its inputs, so it does not establish synthesis. Missing closing tags can still admit a body, and a failed answer score does not roll it back. These limits preserve the implemented trace-fed learning route without upgrading its semantic reliability (RT-RTE-cumulative-update, EPI-RTE-curator-review, EPI-RTE-sheet-admission, EPI-ABS-score-feedback).
Read-back separates requested file restoration from automatic supply to solving contexts. Push selection can be coarse whole-body inclusion, vector-based pair selection, or model judgment about what survives into a replacement sheet. These occur at different operations in a chain. A requested file path or container destination alone does not establish identifier-based push of retained parts (RT-RTE-benchmark-cycle, RT-RTE-context-selection, MEM-OBJ-container-id).
Every axis retains hosted-state uncertainty rather than treating returned text as a description of the consumed payload. Initial content approval and original vector production are also unresolved. The exposed local positives therefore yield partial inventories, not complete-value claims or bounded negatives. The profile asserts no observed operation or causal benefit: no historical result payload was assessed by the supplied reports, and provider internals remain opaque (MEM-OBJ-container-id, MEM-OBJ-input-embeddings, RT-CLM-adaptive-memory, EPI-CLM-performance-gains).
Source-read log: none. Classification used only the supplied accepted members and reconciliation; no source ambiguity required a source read, and no records or evidence were added.