AutoSci memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies accepted records, not new source evidence. No source files were read. No records are declared. Reconciliation supplies no supersessions: broad page/route containers and their narrower epistemic and memory parts remain distinct canonical records.
The retained boundary is not the enclosing model's storage. MEM-OBJ-wiki-pages, MEM-OBJ-ledger and RT-RTE-review-call establish file-backed knowledge, feedback history and volatile review turns respectively. External model parameters and host compaction are excluded from this memory account, not classified as absent. Likewise, installed language-selected skills are static configuration rather than accumulated experience. Graph JSONL remains a file substrate; symbolic graph access and natural-language evidence are separate operative parts of MEM-OBJ-edges.
Repository retention is conditional. RT-RTE-parallel-ingest and RT-RTE-daily-intake retain the ignored-payload/unforced-staging limitation. Their Git routes do not prove that prepared sources reach workers, generated output returns, or every local file survives CI. The memory annotation's broader task-isolation wording is not used to strengthen transport. This preserves the reconciliation's unresolved disagreement without reopening verification.
The other recorded disagreement concerns general versus purpose-specific context packs. MEM-RTE-context-compilation supports consolidation and selection with explicit section budgets. The profile does not infer that a general pack contains experiments or stale entities. Both positive mechanisms and omission limits survive the runtime's less specific description of RT-OBJ-context-brief.
Authority is attached to actual consumers. Knowledge at answering agents, instruction-like failed-idea avoidance, ranking inputs, recovery routing and executable contract/budget checks are different effects. The experiment-results-to-failure-guidance update has learning authority at its update consumer; the resulting reason supplies knowledge or instruction downstream. These findings rely on MEM-RTE-negative-memory, EPI-RTE-verdict-integrate and RT-BAP-retained-context, not on a blanket implication from persistence to learning. Prompt gates remain claimed; browser/direct CLI callers do not authenticate human review.
Read-back distinguishes requested retrieval from automatic context supply. RT-RTE-knowledge-cycle requests files and relevant full pages. A requested rebuild's selector does not make its return push. In contrast, the annotation on RT-RTE-daily-intake establishes automatic newest-card/log-tail preparation for recommendation. Its coarse selection is distinct from lexical scoring of new external papers. The profile does not manufacture targeted identifier push from requested filenames, or infer a complete push inventory from one inspected route.
Trace-learning positives name durable outputs and later consumers. Automatic experimental failure-memory updates and their retained brief projection are afforded routes; ledger feedback and durable bout reconciliation feeding later execution are wired in MEM-RTE-ledger-replay. Raw logging alone is insufficient. MEM-OBJ-experience and EPI-RTE-hiera-summary expose a reporting revision, not a durable full experience dictionary driving the scheduler. EPI-RTE-hiera-select identifies the actual score/bout/tried-config consumers. These distinctions permit architectural trace-learning findings without asserting improved capacity, scientific confirmation, observed activation or causal benefit.
Partial axes retain unresolved initial provenance/admission, arbitrary payload semantics, mixed trace origins and incompletely inspected regeneration, proposal and dissemination routes. No bounded absence records were declared, so this profile does not turn missing recorded facts into absent operations or a system-wide trace-learning negative. Local known units do not erase those alternatives.