IWE memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The comparison concerns reusable retained content and access structures, not every runtime file. The workspace graph is both an in-memory representation and a graph access structure over file-backed documents. It is not evidence of a separate graph database. The serialized proposal/rejection store is outside that graph. See RT-OBJ-graph, MEM-OBJ-documents and RT-OBJ-session-record.
Canonical citations follow the reconciliation's two supersessions: the epistemic proposal and policy-body declarations resolve to MEM-OBJ-proposal and MEM-OBJ-policy-body. These amendments change no classification. The narrower injection mapping remains distinct from the broader MEM-OBJ-policy-knobs; it is not treated as a duplicate or an opaque alternative.
Staging is the consequential automatic admission witness. The host is instructed to extract conversation evidence and software admits proposals without a human selection test. Durable candidates then return through later inbox invocations. That chain supports afforded trace learning even if no final document is selected: a retained continuation context can qualify without successful improvement or a new scientific claim. Final selected document writes and confirmed policy edits instead have manual admission at afforded strength. Code persistence is not an independent automatic semantic extractor. The unresolved generic caller paths preclude an inventory-wide negative or complete learning assessment. See RT-RTE-session-capture, MEM-RTE-inbox-read and WF-RTE-inbox-selection.
Rejection feedback can influence a durable policy revision and thus has learning authority at that update consumer. That does not make the user-confirmed revision an automatic trace-learning write. The edited policy's later force is instruction at the external interpreter and routing at the symbolic settings consumer. Neither policy loading nor rejection counts establish actual criticism, genuine user feedback, or improved capacity. See MEM-OBJ-rejected and RT-RTE-policy-maintenance.
Curation values classify requested host transformations at afforded strength. Semantic compression into standing facts supports consolidation without asserting entailed derivation or novel synthesis. Normalization, rename and graph cleanup alone do not establish those semantic operations. Uninspected lower-level edits remain included and unresolved. See MEM-RTE-graph-restructure and EPI-RTE-link-cleanup.
Read direction and signal are known for the recorded routes even though deployed host delivery is not observed. Session-start selection is coarse push, conditional on invocation and nonempty output groups; it emits metadata and a fixed policy section, not all document bodies. Explicit keys and lexical search serve pull requests, not identifier or inferred-lexical push. The grouped inbox also omits candidate evidence, which requires a full session inbox request. See RT-RTE-session-index, RT-RTE-document-retrieval and MEM-RTE-inbox-read.
No source files were read for this classification. All warrants come from the supplied accepted records and their annotations. Unknown deployed content, initial provenance/control, lower-level maintenance semantics and host receipt remain beside the affected assessments; no observation or causal benefit is claimed.