SkillWeaver memory profile

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

Comparison rationale

This profile classifies accepted records, not source files or observed runs. The library is executable Python plus a separate readable documentation channel. Checkpoint files do not make the store a repository or vector database; dense vectors in the optional process cache supply parametric access representations, not updated provider weights (MEM-OBJ-code, MEM-OBJ-docstring, MEM-RTE-embedding-selection).

Reconciliation supersedes the epistemic skill-code, skill-docstring, function-metadata and semantic-knowledge object declarations with their memory counterparts. All object citations here use those canonical memory declarations. The selected episode and success/benchmark results remain distinct parts of the larger trace/diagnostic inventory; their local routes retain their own evidence strength. In particular, local score retention wiring does not upgrade the defective enclosing benchmark invocation (MEM-OBJ-traces, EPI-RTE-score-retention, RT-RTE-benchmark-evaluation).

The trace-learning positives concern durable behavior-shaping writes, not demonstrated improvement. Synthesis learns into code/documentation; the test shortcut instead changes persistent selection metadata. A syntactic call in a dead branch can pass that shortcut without executing the function. Later policy force remains real despite this evidential weakness (RT-RTE-skill-admission, EPI-RTE-test-check, EPI-RTE-tested-label). Original browser experience supports trajectories and tool traces, not a further event-stream or session-log value merely because metadata stores events or models generated reasoning.

Existing source and experience have learning authority at synthesis; readable descriptions have knowledge and ranking authority at task consumers. Direct executable effects do not by themselves prove one of the controlled authority values, so that consumer remains locally not determinable rather than receiving an invented execution value. Stored reasons are not automatically future guidance: synthesis diagnostics are not reloaded, while documentation may preserve reasons only if the model actually writes them (MEM-OBJ-code, MEM-OBJ-docstring, MEM-OBJ-traces).

Same-name replacement establishes evolution. Selective serialization establishes consolidation even with raw files retained. Prompted generalization does not establish a novel claim; the generated content relation is indeterminate without candidate/input instances. These limits prevent a synthesis finding without suppressing known evolution (EPI-RTE-trajectory-format, EPI-RTE-skill-merge).

The supported read-back findings are pull for established native, builder, adapter and benchmark consumer chains; direction coverage remains partial. Model relevance selection within a consumer request is not independent push, and a function name or checkpoint prefix is not a push signal. Optional embedding selection, schema exports and artifact reconstruction establish capabilities without a recorded later consumer invocation. Their request/return shape therefore establishes neither pull nor push. Push signals are locally inapplicable for the established pull chains, but the unresolved alternatives prevent axis-wide inapplicability; the signal axis is not determinable. External host memory handling remains excluded explicitly, not resolved by the native route (RT-RTE-native-task, RT-RTE-explore-practice, RT-RTE-browser-use, RT-RTE-benchmark-evaluation, MEM-RTE-embedding-selection, MEM-OBJ-docstring, MEM-OBJ-traces).

Unknown initial content provenance and caller control prevent complete lineage, write-agency and trace-route coverage. No manual admission is inferred from starting a workflow or choosing a file. Images and arbitrary semantic payloads prevent a complete form classification. No absence of learning is inferred from an unchanged semantic field, and no observed operation or causal benefit is inferred from wiring (MEM-OBJ-semantic-knowledge, MEM-OBJ-metadata, MEM-OBJ-traces).

Source-read log: none. No frozen-source ambiguity required a direct source read; all warrants remain in the supplied accepted records.