SkillNote memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The comparison concerns accumulated procedures and their supporting state, not all runtime storage. The database, independent release archives, installed copies and staged snapshots have different consumers and rollback boundaries. In particular, an archive can contain scripts and assets whose form and authority cannot inherit the classification of displayed Markdown. See MEM-OBJ-skill, MEM-OBJ-content-version and MEM-OBJ-skill-version.
The reconciliation supersedes EPI-RTE-install-maintenance with MEM-RTE-install-maintenance. All profile maintenance citations use that canonical record. This affects file replacement, manifest routing, automatic admission, evolution and requested read-back; it adds neither atomic rollback nor universal withdrawal. Other overlapping text, feedback and endpoint records remain valid containers or functional parts, not superseded identities.
Instruction delivery is implemented, while its force at the external model is an afforded instruction path. The foreground OpenClaw comparison of descriptions with a task is requested model judgment, not a wired semantic ranker. By contrast, retained comment ratings and usage counts influence code-level relative priority in the requested bundle. That is ranking authority, even though the response is transient and the scoring policy is fixed. The comment prefix supplies reference or advice to the documented consumer, not verified task truth. See RT-BAP-host-instructions, RT-RTE-openclaw-use and MEM-RTE-feedback-catalog.
The automatic sync chain has both a requested registry fetch and automatic supply toward later host discovery. Collection membership actually selects delivered parts, supporting identifier-based push; wildcard availability is coarse. Neither a requested MCP slug nor a task-conditioned foreground read is itself a push signal. PostCompact output does not become confirmed model-visible delivery merely because a catalog was fetched. See RT-RTE-project-sync, RT-RTE-mcp-delivery and RT-RTE-context-hooks.
Curation is local. A staged body promoted to current content gains a consumption tier. Disabled archives retain history while losing covered download eligibility. A 30-day usage window removes old reports' ranking influence, not their stored rows. Deduplicating a repeated rating updates an assessment, not a procedure. These mechanisms do not imply semantic improvement, global revocation or a synthesized procedure. See MEM-RTE-inbound-resolution, MEM-OBJ-skill-version, EPI-RTE-bundle-metadata and EPI-RTE-rating-admission.
Trace extraction is not automatically trace learning. The watcher retains analytics events and replay state, not revised guidance; its events are not the usage store counted by the bundle. The latter recomputes a candidate payload from retained reports without persisting a derived procedure or learned ranking artifact. This supports trace-extracted lineage and feedback-dependent ranking, not a qualifying positive learning finding. Uninspected evaluation consumers and unknown caller/payload production also prevent a complete negative. Therefore trace source remains unresolved rather than being inferred from the visible JSONL input or declared inapplicable. See MEM-RTE-trace-reporting, MEM-RTE-feedback-catalog and MEM-OBJ-skill-version.
No frozen-source files were read for this profile. Classification uses only the supplied accepted records and their annotations, with reconciliation resolving identity. No operation or benefit is upgraded to observed or causally supported.