llm-wiki memory profile

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

Comparison rationale

This profile classifies accepted records, not fresh source evidence. No source files were read. Wiring supports persistence, selectors and delivery, not observed compliance or benefit.

Raw source and conditional excerpt are separate units. MEM-OBJ-reference-source is the canonical raw asset after reconciliation; EPI-OBJ-derived-reference remains the conditional concept part. A readable copied Markdown source can skip generation of the latter. Arbitrary resource encoding and initial provenance therefore remain unresolved rather than inheriting the excerpt's readable display.

Reconciliation also makes MEM-RTE-directory-deprecation and MEM-RTE-index-refresh canonical for the duplicate retirement and refresh routes. Soft withdrawal supports invalidate even though search/get still deliver the body. Index regeneration is not consolidation or synthesis; standalone refresh does not imply a change-log write. Neither operation establishes acceptance of domain claims.

Retained text is a ranking input at lexical search and knowledge at the answering host. Editable rule prose supplies instruction, while the parsed AGENTS type list supplies validation and an enforced create veto. These do not make optional diagnostics a retained universal admission policy. Confirmation flags establish neither an authenticated human decision nor fresh approval of revised bodies (RT-RTE-concept-create, RT-RTE-concept-update).

The lifecycle push is broad, budgeted rule collection with a coarse signal. Requested search/get content and explicit rules calls remain pull. CLI get/create/update additionally push a scoped rule appendix without a request for those rule files. Its target-directory ancestor match selects actual retained APPEND files and AGENTS gate sections, supporting identifier selection independently of the requested concept return. The appendix omits other AGENTS prose and may silently fail; these limits do not turn it into a requested rules return. A session-keyed digest only suppresses unchanged lifecycle text and does not itself establish identifier-based targeting (RT-RTE-context-delivery, RT-RTE-bundle-read).

The records retain a caller-mediated experience-recording affordance, not a specified recorded-trace-to-derived-memory producer. EPI-CLM-knowledge-recording requests investigation and recording; it does not name the original recorded inputs consumed by the external semantic producer. Operation logs and counters alone do not supply that missing chain. No member declares an evidenced absence, so neither trace-learning no nor trace-source inapplicability is warranted. These limits do not deny persistence, the recording affordance or possible host learning.