Hyperagents memory profile

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

Comparison rationale

This revision-2 profile classifies accepted records, not source files or an observed run. All positive findings retain the basis of the actual route. No observed or causal basis is available. Each axis is partial because its explicit inventory retains unresolved alternatives rather than promoting a local witness into complete system coverage.

Retained carriers and operative parts

RT-OBJ-model-patch and RT-BAP-code-replay establish file retention and repository reconstruction. Symbolic executable directions have instruction authority at Python execution; they are not merely advisory model text. This mapping does not assign every conceivable effect of generated code. Reward code is a separate executable part, and checkpoint parameters are a separate parametric part. The checkpoint file establishes files, not a model-weights storage backend. Its opaque auxiliary payload stays unresolved under RT-CMP-policy-runner.

The prediction CSV is not counted as identical to its answer cells. The broader saved-report object is not identical to its label-agreement, proof-grade or environment-summary parts. These distinctions preserve the separate consumers in MEM-OBJ-predictions, MEM-OBJ-report, EPI-OBJ-proof-grades and EPI-RTE-outcome-summary. Reconciliation supplies no supersessions affecting any value; all citations therefore retain their declaring identities. Overlap is not duplicate evidence of a separate mechanism.

Admission, authority and maintenance

Automatic describes admission without a per-content operator decision, not whether a model authored the proposal or a human started the workflow. Original bootstrap approval and separately supplied checkpoint admission remain unknown even though subsequent software updates are known. Metadata criteria can validate and veto reuse while the same IDs and paths route execution. Scores rank candidates and gate continuation without certifying improvement; starting-node exceptions and fallback remain part of that scope. See RT-BAP-selector-flags and RT-RTE-parent-selection.

Evolution revises inherited code. Invalidation withdraws reliance while preserving archive history. DGM field extraction consolidates the logged diagnosis into a reduced implementation statement, but its diagnosis-generation semantics do not establish synthesis. Copy pruning changes another invocation's view, not the retained host archive. These are distinct operations in RT-RTE-editor-admission, MEM-RTE-dgm-diagnosis and MEM-RTE-archive-delivery.

Selection and trace-fed updates

A foreground file request is pull. Automatic archive exposure and selected lineage replay are push operations at a different stage. The normal copy's coarse filename filtering is not a relevance selector. The selected-parent identity match retrieves actual patch parts; prediction reuse similarly matches current question IDs to retained answers. Conversely, a filename on a requested checkpoint or the question-ID lookup fulfilling an ensemble answer request does not independently establish identifier-based push. See MEM-RTE-prediction-reuse and RT-RTE-optional-consumers.

The DGM write uses retained experience to generate logged guidance consumed by another editor. Its update consumer gives traces learning authority; its output supplies instruction authority. Genesis separately wires rollout-fed numerical adaptation and checkpoint inference. These qualify for trace_learning without observed improvement, accepted explanatory claims or a theory-builder finding. Ordinary archive-informed editing remains afforded: copied evidence is not automatically read by the model. Raw-log retention and score-based reuse alone establish no additional qualifying update.

Session turns, recorded tool/execution feedback and linked episode experience support different trace-source findings. Static task/reference inputs do not become traces by being embedded in a diagnosis prompt. No event-streams value is inferred solely from episode timestamps, CSV transport or a broad log label. The DGM retry defect limits recovery, not the supported initial branch. MEM-RTE-dgm-diagnosis and RT-CMP-policy-runner retain these warrants.

Coverage limits and source-read log

Unknown initial provenance prevents complete lineage and admission coverage. Opaque checkpoint fields, provider retention, full environment-context derivation and detailed transfer consumers prevent complete enumeration of forms, authorities, transformations, directions and trace origins. Images are not silently classified by their filenames. A known local route does not turn these alternatives into absence or inapplicability. Unchanged shipped guidance and fixed inference weights are excluded as such because the scope requires through-use retention; possible provider-side writes remain explicitly included as uninspected in RT-CMP-provider-models.

No source files were read for this classification. The supplied accepted members and reconciliation sufficed; no records, quotations, evidence sources or verification judgments were added.