ModularRSI memory profile

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

Comparison rationale

The principal accumulated behavior artifact is the executable generation, not the optional editor-history display. The library is file-backed symbolic strategy, while its registration descriptions are language interpreted by a separate composer. Findings and proposals retain reasons upstream; the solver does not automatically receive those reasons. These distinctions support different form, lineage and consumer-authority findings rather than one classification for everything stored in Python or JSON. See RT-OBJ-module-library, RT-OBJ-module-description, MEM-RTE-evidence-brief and MEM-OBJ-proposal-record.

All findings remain at their recorded evidence layers. In particular, causal synthesis is afforded by investigator instructions; no actual newly formulated diagnosis was inspected. Consolidation describes the reduction route, not a fidelity guarantee. The epistemic account's indeterminate semantic relation for generated summaries prevents claiming that every summary is only reshaping; it does not remove the wired compression operation. Same-intervention linking is deduplication of proposed work through support merging, not erasure of raw findings. See EPI-OBJ-contrast-finding, RT-RTE-context-handoff and EPI-RTE-finding-disposition.

Trace learning uses this profile's automatic trace-fed write criterion, not demonstrated future improvement. Diagnosis-to-code, continuation summarization and outcome-fed archive retirement have qualifying later consumers. Their experience can have learning authority at an evidenced update consumer while output code, guidance or eligibility state has a different downstream force. In particular, QA continuation has knowledge and instruction authority at the next solver inference call, not learning authority: that call is not evidenced to update parameters or another durable behavior-shaping artifact. Wiring does not establish that a particular change was exercised or beneficial. Original source classes follow the content actually consumed: conversational turns, tool feedback and linked multi-step task experience, not the fact that ledger rows are serialized as events. The primary diagnosis route supports trajectories independently; confirmation's task/bundle/reward tuples do not establish that source. Confirmation's trace-source unit is therefore not-determinable, without withdrawing its qualifying automatic archive update or borrowing the diagnosis route's richer input. See MEM-RTE-evidence-brief, RT-RTE-context-handoff, EPI-RTE-paired-confirmation and RT-RTE-generation-change.

Read-back distinguishes driver-selected pushes from editor-requested returns. An exact task/finding/cache match establishes identifier selection only where it actually selects a delivered retained part; a path on a requested file is not an extra push signal. Model catalog judgment is a separate selector from coarse archive eligibility and budget clipping. Static skill catalogs are outside this accumulated-memory inventory, so the uninspected optional solver skill-injection path does not become a finding about learned instructions. See MEM-RTE-history-contrast, MEM-RTE-evidence-brief, RT-RTE-bundle-selection and RT-RTE-solver-task.

Reconciliation supplies no supersessions; no value changes identity through an amendment. Its retained editor-memory ambiguity prevents concluding automatic journal-to-editor delivery. The profile uses the bounded absence of searched caller wiring and separately preserves positive journal retention, rather than treating a formatter as a live consumer. It also keeps manifest and cross-module receiving routes unresolved. Retryable findings are not classified as scheduled future work merely because they remain nonterminal; already-open proposal reuse is independently wired. See MEM-ABS-editor-readback, MEM-OBJ-editor-memory, MEM-OBJ-implementation-record and EPI-RTE-finding-disposition.

Each axis is partial because the included unresolved parts cannot be erased by an inspected baseline witness. Missing initial training/admission history limits lineage, write agency and trace origins; uninspected strategy bodies limit exhaustive mechanism coverage; provider persistence remains opaque. Static model weights are not assigned parametric memory, model-weights storage or learning merely because inference uses them. The searched journal-feedback path has a bounded local trace-learning no, but there is no complete learning-inventory negative and no supported system-wide no. Derived unions may retain the positive findings and their bases, but cannot turn them into complete-value statistics. See MEM-OBJ-archive-entry, RT-OBJ-module-library, RT-RTE-context-handoff and RT-CMP-lite-llm.

Source-read log: none. This classification uses only the supplied accepted members and their annotations; no frozen-source files were newly read, no evidence was added and report verification was not reopened.