SkillLift memory profile

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

Comparison rationale

The comparison covers retained improvement-plane memory, not every object in SkillLift's runtime. File trees, patch recipes, JSON checkpoints, receipt rules and recorded outcome history reach later rounds or test invocations. The ephemeral evidence view is a transformation of these inputs, not an additional memory store. Remote inference components do not establish use-accumulated provider memory: RT-CMP-llm-client leaves provider mutability unresolved. This profile does not turn that runtime component's distributed-parametric form into a memory finding.

Natural-language bodies and symbolic control fields have different consumers. Retained skill text supplies instructions to benchmark agents; retained hashes, accepted IDs and validity fields instead route or veto reconstruction and reuse. Receipt text directs verifier judgments, while points and hits determine ranking and code thresholds. Retained score/proposal experience also has learning authority at the update consumer, independently of the generated receipt's instruction authority at the next verifier. See MEM-OBJ-candidate-record, MEM-OBJ-search-state, MEM-OBJ-receipt-rubrics and RT-BAP-receipt-scoring. None of these classifications claims observed compliance, accurate diagnosis or demonstrated improvement.

The qualifying trace-learning witnesses are automatic durable updates from recorded evaluation/proposal events to later search, skill or scoring behavior. Full benchmark conversations are not needed for these witnesses. Their original recorded input is discrete runtime outcome experience, hence event-streams, not a session-log merely because models participate. Static task text and skill bodies remain nontrace inputs. The message/tool-derived summary in MEM-OBJ-behavior-summary is different: it has trace-extracted lineage but its feedback metadata is projected out of shared receipt revision. Saved raw trajectories and uninspected readers cannot establish either additional trace-learning sources or a global negative.

Read-back in the assessed chains is pull. Coordinator requests for planning, replay, calibration and testing cause code to fetch the specified material. A selector model choosing optional editor files fulfills that request; it is not independent push. Thus IDs and budgets on these chains supply no positive push-signal finding. Unresolved consumers leave the system-level signal axis uninspected rather than inapplicable. See RT-RTE-branch-refinement, RT-RTE-task-evolution and RT-RTE-domain-fusion.

Wired evolution and promotion do not certify semantic curation. Domain fusion requests deduplication and preservation of disputed rules; category fusion requests complementary workflows. Their structural checks do not prove these transformations occurred, so consolidation and deduplication keep claimed bases. No linked output establishes synthesis of a new claim. Checkpoint cleanup and materialized-tree replacement are not by themselves decay or history-preserving invalidation. See MEM-OBJ-master-skill, EPI-RTE-category-fusion and MEM-OBJ-cluster-checkpoint.

Reconciliation establishes no supersessions, so each citation retains its canonical declaring member. Its corrected domain connection matters to lineage and selection: RT-RTE-domain-fusion exports/loads every current cluster skill, not just an oracle-ranked winner. Category fusion still uses batch winners. Its corrected evidence-packing connection also matters: exact SKILL.md key protection does not cover all prefixed instruction paths under a budget, while ordinary task Mode B uses no evidence budget and can reject an oversized request instead. See EPI-RTE-evidence-packing and the memory annotation on RT-RTE-receipt-revision. Neither corrected connection upgrades semantic checking.

The conditional replay defect in RT-CLM-refinement-replay qualifies persistence of accepted refined task bodies, not existence of the revision route or category/domain full snapshots. Initial content control, auxiliary payload consumers, raw trajectory readers and other shipped baseline/entry-point mechanisms lack complete accepted classification records. They remain included unresolved alternatives; positive witnesses cannot make the inventory complete. No absence records justify system-wide negative values.

Source-read log: none. The supplied accepted records, their annotations and reconciliation resolved the classification questions without frozen-source inspection. No records, evidence passages or source identities were added.