PrimeScientist memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The comparison covers retained experience and its later consumers, not every file or model used by the workflow. All positive bases are implementation, requested behavior or affordance, not observed operation. File staging does not establish native-skill loading, model compliance, scientific warrant or improved future capacity. The symbolic Q/visits route and automatic diagnostic reductions nevertheless meet the profile's trace-learning test: they update retained behavior-shaping state with subsequent consumers. See RT-OBJ-stats, RT-BAP-score-selector, MEM-RTE-log-brief, and MEM-OBJ-history-shown.
The file substrate does not turn every operative part into language. Numeric priors/statistics, identifiers and executable code have symbolic consumers; reasons and diagnostic text have interpretation paths. Arbitrary workspace model/data files remain uninspected rather than being classified as parametric memory. Provider weights are outside the accumulated-memory boundary because no recorded update route brings them inside it. Session payload remains included and opaque, independent of the inspected JSON access metadata. See MEM-OBJ-prior, RT-OBJ-sandbox, and MEM-RTE-session-resume.
Known copying of insight files establishes imported lineage and automatic supply, not initial authorship, human approval, experience extraction or a closed cross-task learning loop. The current-task name match and nonempty file filter select delivered parts without inferred relevance. Initial insight admission remains unresolved. Existing raw proposal directories have the same initial-control limitation even though subsequent bootstrap is automatic. See MEM-OBJ-insights, MEM-RTE-insight-delivery, and RT-RTE-plan-admission.
Head/tail excerpts, short failure digests and bounded history consolidate retained experience. No output claim proves synthesis by the reflector: generated proposal semantics are unavailable. Score hiding instead establishes withdrawal of normal branch reliance while retaining history. Strict score improvement replaces baseline code; it is not acceptance of an explanatory claim. Trace-source coverage retains unresolved narrative and proposal inputs rather than assigning event streams or session logs from filenames. See MEM-OBJ-pruned-children, RT-RTE-vanilla-retention, and RT-RTE-plan-admission.
The reconciliation supersedes EPI-OBJ-prior, EPI-OBJ-log-brief and EPI-OBJ-packet with their MEM declarations. All profile references use the canonical memory records; the supersessions affect reason/form/lineage, diagnostic reduction and packet-consumer classifications, not their producing route identities. The failed-resume recovery disagreement is retained as a limitation: neither unconditional clearing nor a guaranteed fresh next context is asserted here. The visible resume request supports a route, not restoration fidelity. See MEM-OBJ-prior, MEM-OBJ-log-brief, MEM-OBJ-packet, RT-RTE-plan-admission, and MEM-RTE-session-resume.
FIRE-Bench packet consumers receive normalized scores but not reference-derived false-positive/negative diagnoses. Retention is therefore not complete diagnostic read-back or reliable pointer dereference after pruning. Replay can also lose conclusion-only narrative while retaining scores/actions. These limits qualify knowledge and trace-source coverage without erasing numeric control or the positive history route. See EPI-RTE-score-attachment, MEM-OBJ-packet, and MEM-OBJ-history-shown.
Source-read log: none. Classification used only the supplied accepted members and reconciliation; no ambiguity required a frozen-source read. Every axis remains partial because its inventory retains unresolved included parts. No system-wide negative or complete-value enumeration is inferred from a known branch.