Prove2Me_workspace memory profile

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

Comparison rationale

The memory account inventories service objects and filesystem retention, not an implemented backend database. Dependency graphs describe relations among formal objects; they do not establish a graph storage substrate. The same retained object may have prose and symbolic parts, so those parts have separate form and consumer units. See MEM-OBJ-library-declarations, MEM-OBJ-workspace-files and MEM-OBJ-upload-state.

All supported mechanisms here are claimed unless a finding explicitly cites wired extraction. That implementation witness covers fact production only. It does not upgrade the unshipped converter, recovery bookkeeping, service persistence, context isolation or actual later use. See EPI-RTE-extraction-facts and RT-RTE-project-upload.

Reconciliation supplies no supersessions. Containers and their operative parts remain canonical: formal targets and definitions are parts of the broader library account, while extraction facts are distinct from the wider upload state. This profile therefore cites the record appropriate to each local claim without treating overlaps as additional mechanisms.

Behavioral authority follows consumers. Testimony and past attempts supply knowledge; captain specifications direct later solver behavior; saved pointers and recovery IDs route operations; votes, ordering and campaign data rank results; retained templates/specifications validate matching candidates. These forces do not certify mathematical truth. No learning authority is inferred from merely retaining criticism. See MEM-BAP-experience-guidance, RT-BAP-milestone-guidance, EPI-RTE-quality-signals, EPI-RTE-campaign-review and EPI-RTE-campaign-selection.

Admission is local to the write. An agent can directly admit a proof, comment or mirror without per-content human approval, while proposal launch and campaign attestation explicitly require human decisions. Generic account/editor identity cannot classify all remaining writes. A human starting the workflow does not make its subsequent admissions manual. See RT-RTE-solver-proof, MEM-RTE-discussion-feedback, RT-RTE-proposal-audit and EPI-RTE-campaign-admission.

Requested HTTP retrieval, file loading and campaign browsing are pull even when the service filters or orders the answer. There is no separately established unsolicited selection within those chains. The absence of a recorded host injection mechanism is not bounded evidence of its absence. See MEM-RTE-solver-scouting and RT-CMP-agent-model.

Experience commentary has a later reader, but the accepted records stop short of identifying the recorded trace inputs actually consumed by an automatic update. The unresolved trace axes do not deny a possible agent-authored trace-learning route, demand model-weight changes, or demand demonstrated improvement. Nor does retaining raw failed code establish such a route. See MEM-RTE-discussion-feedback and MEM-OBJ-submission-explanation.

No direct source reads were made. Classification uses only the supplied accepted records and their annotations; no new records, evidence or quotations are introduced. Missing backend and host implementation, live-instance provenance and specific later consumers remain local limits rather than global negative findings.