Kompl memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies accepted records, not source files or a live instance. All positive implementation findings are local wiring or affordances; none are observations of operation, compliance, grounding or benefit. The local retained-object inventory comes from MEM-OBJ-raw, MEM-OBJ-extractions, MEM-OBJ-pages, MEM-OBJ-wiki-pages and their separate access, proposal and conversation records. Encoding does not determine storage substrate. Dense embeddings are parametric representations in vector storage, not updated MiniLM weights. Restored embeddings retain an imported path without resolved pre-import provenance.
The reconciliation supersedes the epistemic manual-save, lint and saved-links routes with MEM-RTE-manual-save, MEM-RTE-lint-retention and MEM-RTE-saved-links. Classification cites those canonical destinations. In particular, the saved-links route retains the complementary empty-set witness: dismissal can produce an imported-all assertion not entailed by the retained failures. That implemented deterministic branch supports synthesis independently of model prose whose actual semantic transformation is unknown. The extraction bundle and its model-output part, and the plan control container and its prose payload, remain distinct valid referents; they are not superseded wholesale.
Read-back distinguishes upstream request from downstream augmentation. RT-RTE-mcp-read is requested delivery. Question-driven page retrieval in RT-RTE-chat-feedback also remains pull. Automatically supplying saved conventions, target-page prose and earlier-layer draft summaries to a drafting model is a separate context-augmentation operation on RT-RTE-plan-draft; the model makes no memory request. Likewise, code adds old conversation turns to answer synthesis under MEM-RTE-history-read. Identity, relevance and coarse budget selectors are retained separately. Requested query ranking is not converted into unsolicited push merely because code selects and delivers pages.
The runtime and memory mechanisms now agree on earliest-row history selection and ordinary FAQ provenance failure. The reconciliation retains conflicts only about memory's characterization of an earlier runtime position. They do not prevent these bounded classifications: MEM-RTE-history-read establishes replay selection without summary production; MEM-CLM-chat-provenance prevents treating ordinary pending FAQ creation as a universally wired wiki-compounding cycle. Empty-citation manual saves and coincident page/source IDs are exceptions, not repairs of ordinary automatic adoption.
Trace learning is assessed at automatic durable trace-fed writes, not at raw log storage or parameter training. MEM-RTE-saved-links has a concrete trace-to-page-to-later-consumer route. RT-RTE-chat-feedback has an automatic trace-to-pending-candidate route with a later reviewer, while ordinary wiki installation is constrained. Learning authority at those update consumers and knowledge authority at the downstream page/candidate readers are separate paths. Neither establishes improved capacity or exercised self-improvement. Generic caller identity does not prove manual admission; the UI's explicit content-selection/review route affords it without resolving arbitrary API callers.
Every axis remains partial because local positives do not exhaust hidden continuation, supplied replacement provenance and other named included alternatives. Static weights are excluded from accumulated memory, not declared absent learning. No absence records warrant a global negative for trace learning, provider memory, or read-back. Generated schema rationale, actual paired model transformations, original imported embedding derivation and hidden checkpoint inputs are missing recorded facts; no source understanding is substituted for them.
Source-read log: none. This worker read only the supplied invocation, contracts and pinned members; it performed no additional target-source reads.