jaz-evals memory profile

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

Comparison rationale

All ten axes are partial. The accepted records establish real advice delivery and diagnostic read-back, but do not enumerate external Letta memory internals or unaudited publication consumers. A known wired ACE witness cannot resolve those alternatives. The profile classifies implementation and afforded model-mediated routes, not exercised learning or improved outcomes.

ACE has distinct operative parts. Advice is interpreted natural language; counts and IDs are symbolic. Counts do not establish reliability weighting or ranking. Advice bodies do influence similarity grouping, so they have ranking authority at that consumer. Recorded trajectories have learning authority at the update consumer, while the resulting advisory playbook has knowledge authority at the solver. This does not make advice mandatory instruction or make static parser rules retained validation memory. See MEM-OBJ-playbook, MEM-OBJ-playbook-counts, MEM-RTE-playbook-dedup and RT-RTE-ace-adaptation.

Similarity refinement supports deduplication and replacement, including the merge-free branch. Model merging affords consolidation but establishes neither semantic fidelity nor synthesis of a new claim. Trace-conditioned ADDs likewise do not settle the semantic relation of particular advice to its inputs. No generated example supports a stronger transformation classification. See EPI-RTE-dedup-replacement and MEM-RTE-playbook-dedup.

ACE prompt supply is coarse push: all current advice is inserted automatically. Optional seed loading, foreground history searches and operator analysis fulfill explicit requests. Identifier-bearing paths and substring searches on those pull operations are not identifier or lexical push. JAZ root-authored revisions have a separate afforded judgment-selected push to subsequent delegates. See MEM-OBJ-playbook, MEM-RTE-jaz-history, MEM-RTE-smolagents-history and RT-RTE-jaz-queue.

Requested continuation summaries qualify locally as trace learning because model-authored trace-derived context is retained for a successor, even without a new claim or demonstrated improvement. The smolagents wrapper display suffix does not classify the complete payload as a summary. Letta's inspected summary return does not establish its retention or injection; that route remains unresolved, not negative. Optional dense archive representations have only the adapter's descriptive warrant, not a wired vector-store classification. See MEM-RTE-smolagents-history, MEM-OBJ-compaction-summary and MEM-OBJ-archival-memory.

The reconciliation supersedes EPI-RTE-analysis-report with MEM-RTE-artifact-analysis; all diagnostic classifications use the latter. Advice/count objects and the combined playbook container remain valid separately, as do grouping and replacement subroutes. No supersession strengthens evidence or converts structural admission into epistemic acceptance.

Source-read log: none. Classification used only the supplied accepted members and reconciliation; no source ambiguity required a new read. Missing initialization control prevents a manual-write classification; missing external consumer and installation facts prevent complete authority, curation, read-back and trace-learning assessments. No bounded absence record licenses a system-wide negative.