Wuphf memory profile

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

Comparison rationale

The boundary is a complete artifact, partial loop. Every axis is partial because the accepted inventory retains unresolved material alternatives, not because wired implementation needs a live run to be classifiable. No finding claims observation, causal benefit, or universally complete value coverage. Each positive retains its local route and consumer; the union is not a complete inventory assertion.

Reconciliation identifies six duplicate referents. This profile uses the canonical memory destinations for typed facts, learning records, artifact extraction, entity synthesis, task distillation and learning admission. Their superseded epistemic declarations do not supply stronger statuses. The epistemic entity-brief, verification, lint and playbook-compiler records remain distinct. The knowledge-packet record overlaps only the learning/index branch of MEM-RTE-task-knowledge; article-pointer delivery does not establish article-body delivery. Enclosing office turns likewise remain distinct from their contained selectors.

The most consequential authority split is task experience at its automatic update consumer versus derived text at the later office model. MEM-RTE-task-distillation supports learning force and trace learning in the profile's route sense; MEM-BAP-knowledge-guidance supports downstream reference force. Neither establishes improved capacity. Verification feedback has only the narrow command/file/HTTP license described by EPI-RTE-task-check. Copied task narratives retain unresolved original provenance, even when a pass result gates their retention.

The error-directed repair branch of RT-RTE-tool-authoring also qualifies as wired trace learning: simulated execution feedback feeds automatic repair, admitted code persists, and a later worker consumes it. Conditional activation is not missing wiring. This does not contradict the reports' unestablished learning/improvement outcomes, and it does not turn raw routine transcripts into a feedback loop. Simulation bounds the feedback's business warrant, not its execution-trace origin.

The procedural wrapper supplies an afforded instruction path, not a fully traced invocation. EPI-RTE-playbook-compile and MEM-RTE-skill-registration leave loading/assignment unresolved. Stored JavaScript instead reaches a wired executor through RT-BAP-tool-execution. Its operation selection is executable routing, while rejection tombstones have distinct admission validation and enforcement force. Current-run approval flags and static schemas are not silently included as accumulated memory controls.

Compression, extraction and fact-fed rewriting support consolidation/evolution without proving semantic fidelity. The name synthesis and a model's involvement do not establish a new claim absent from inputs. MEM-RTE-brief-synthesis retains that uncertainty. Archive and contradiction expiry have bounded withdrawal effects, not global deletion of derivatives; MEM-RTE-wiki-maintenance and EPI-RTE-lint-disposal prevent the stronger conclusion.

Read-back directions classify operations within a chain. Automatically querying a backend to fill a notification is push to the recipient model; an explicitly requested article return is pull. The identity signal on MEM-RTE-upstream-results actually selects delivered task records. Remote query ranking stays opaque despite local lexical and dense witnesses. Routine outcomes have a requested frontend presentation path through the workflow annotation on RT-RTE-scheduled-tool, not an established conversational feedback loop into future authoring.

No frozen source files were read for this classification. All warrants come from the supplied accepted records and their annotations; no new evidence, records, quotes or verification judgments were added. Unknown remote storage is not recast as a particular database, a display projection does not classify opaque native payloads, and the current-state projection in MEM-OBJ-session-memory-snapshot is not asserted to be a durable checkpoint. Additional source inspection would require recorded facts before it could strengthen this profile.