ExpeL memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies the accepted records, not a target execution. All positive bases remain local: the shipped reflection and rule routes are wired, while auxiliary helper invocation is only afforded. No supplied execution or comparison upgrades these findings to observed operation or causal benefit (RT-RTE-training-loop, RT-RTE-rule-admission, MEM-RTE-auxiliary-critique).
Storage and form have complete classification within the recorded object inventory. Other axes retain the specific unresolved facts: auxiliary/archive consumers, caller content control, checkpoint provenance, embedding provenance and generated semantic transformations. An unresolved consumer does not make an otherwise described object's substrate or form unknowable. External model weights are components, not the accumulated memory boundary; their opacity is not used to invent model-weight memory or to erase the recorded dense retrieval-access part (MEM-OBJ-retrieval-docs, RT-CMP-embedder).
The reconciliation supersedes the epistemic rule-text, counter and auxiliary-dictionary object declarations with MEM-OBJ-rule-items, MEM-OBJ-rule-counts and MEM-OBJ-critiques. Those canonical objects are cited here. Epistemic proposal, checking, disposition and cleanup parts remain distinct accepted routes, so their local semantic limits and counter effects remain usable. The reconciliation's remaining discrepancy concerns how epistemic describes the current runtime's fold qualification, not the selection mechanism. This profile relies on the conditional mechanism now explicit in RT-RTE-rule-admission, not the characterization of an unconditional peer claim.
Learning authority belongs to retained trials/current rules at the rule-update consumer; instruction and advice belong to derived guidance at later acting consumers. Counter eligibility supplies validation and an enforced membership decision independently of model-interpreted rule content. Positive trace learning names the automatic retained-guidance chains, not successful improvement, accepted knowledge or fine-tuning (EPI-RTE-counter-disposition, RT-BAP-rules-acting, MEM-OBJ-reflections).
Curation distinguishes fixed-code revision, priority changes and withdrawal from requested semantic generalization. A prompt asking for generally applicable future insights supplies only a claimed synthesis finding. Actual generated rule semantics remain indeterminate; the profile does not upgrade that request to wired production of novel claims, or classify every model extraction as consolidation. Substring duplicate-proposal filtering is not evidence of merging/removing duplicate retained entries (EPI-RTE-insight-proposal, EPI-CLM-general-insights, EPI-RTE-syntax-check).
Read-back separates requested saved-state/log loading from subsequent automatic model context supply. Embedding matching, task-identity lookup and coarse eligibility operate at different steps; the matched document can point to a different shortest trajectory from the one that supplied its key. Fold extraction and retrieval have different pools, so the profile makes no complete held-out isolation claim (RT-RTE-checkpoint-restore, MEM-RTE-log-inspection, MEM-OBJ-retrieval-docs, RT-RTE-evaluation-context).
Source-read log: none. Classification used only the supplied accepted members and their records; no source ambiguity required an additional frozen-source read.