OpenSage memory profile

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

Comparison rationale

All ten axes are partial. Positive file/context routes do not close the inventory of service payloads, live retention, trainer updates or extension-specific alternatives. A missing deployed run limits observation and benefit, not the wired summary routes. The profile classifies the accepted records only; no target source was read.

Reconciliation supersedes the epistemic shared-findings object and reduced-response object with MEM-OBJ-findings and MEM-OBJ-tool-summary, and its reduction route with MEM-RTE-output-reduction. Those canonical citations govern the corresponding values. The finding check/write/delivery records remain distinct functions, so their internal wired evidence does not upgrade the afforded shipped invocation of MEM-RTE-finding-admission.

Compaction and tool reduction are consolidation even when originals remain on disk or semantic fidelity fails. The summary-update consumers receive recorded experience; later agents receive derived context. This supports learning authority at the update consumer and knowledge at the later model, without proving improved capacity. The qualifying trace-learning chain is automatic trace-fed production, retention and later invocation, not merely checkpointing or exporting raw logs. MEM-OBJ-compacted-content retains non-cumulative range replacement and unchecked rationale preservation; these limits remain beside the wired positive.

Read-back direction describes operations, not an entire chain. Requested resume/history consultation is pull; later automatic history folding and benchmark prompt inclusion are push. The branch/range selector in MEM-OBJ-compacted-content supports identifier and coarse signals separately. A requested checkpoint path or a saved-output file ID does not independently support targeted push. The binary payload in MEM-OBJ-image-parts remains unclassified by form. The symbolic MIME/part wrapper and natural-language rejection notice occupy separate known units: MEM-RTE-image-retention establishes the notice's interpreted request to resubmit smaller images. Neither unit classifies the binary payload.

Policy-requested wiki revision/removal stays claimed. EPI-RTE-finding-write implements duplicate suppression, not withdrawal of an older contradicted claim or truth acceptance. Retained previous findings supply validation and an enforced admission veto, while broadcast supplies advice. Structural checkpoint parsing and image request filtering do not automatically imply retained validation criteria or curation.

Retained benchmark grades and scores have downstream reporting/ranking consumers, distinct from ordinary agent memory and from trainer learning. WF-RTE-devops-grading reads stored grades into aggregate lists without sending failure back for policy revision. External checkpoint selection and parameter updates remain uninspected in RT-RTE-evaluation-rollout. Static root policy and ordinary provider inference weights are not counted as accumulated memory.

The partial lineage inventory does not enumerate the caller-finding and appended-analysis derivations in MEM-OBJ-findings and EPI-RTE-finding-write, or the model-generation/parsing/serialization path in MEM-OBJ-vulnerability-findings and MEM-RTE-findings-resume. Withhold object-complete lineage comparisons for these parts; the emitted witnesses do not establish that coverage.