KBLaM memory profile

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

Comparison rationale

This profile classifies accepted records, not new source observations. The inventory separates semantic knowledge from its index/field structure, dense base representations, learned machinery and retained experiment evidence. A dense array is a file-backed representation, not evidence of a vector store. Model weights are separately retained parameter state, not prose instructions. See MEM-OBJ-embd-key, MEM-OBJ-embd-value and MEM-OBJ-encoder.

Coverage and strength are independent. Wired local transformations survive missing historical provenance and unobserved operation, but they cannot resolve opaque initialization or provider retention. The unchanged evaluator import failure, generator typo, processor compatibility issues, resume layout mismatch and scorer defects remain limits on reachability and completed operation. No finding claims an observed activation or causal benefit. See RT-RTE-accuracy-generation, RT-RTE-dataset-production, RT-RTE-direct-inference, RT-RTE-adapter-training and MEM-OBJ-open-judgment.

Requested loading and subsequent automatic context selection are different operations. Paths and row indices do not by themselves establish identifier push. Random/subset driver selection is coarse; the optional Llama top-k selector uses current queries and retained numerical keys. The score-slicing qualification prevents treating that pruning as a proven equivalent optimization. See RT-RTE-kb-attention and RT-RTE-accuracy-generation.

Learning authority belongs to QA targets at the optimizer. It does not imply trace learning: the recorded local supervision is supplied/generated content, not an established recorded-experience source. The accepted set declares no bounded-absence records, so neither a controlled trace-learning no nor trace-source inapplicability is asserted. Conversely answer traces and attention snapshots do feed retained analysis, but the recorded downstream consumers do not turn those analyses into durable answering behavior. The two judgment parts remain separate: scalar reference agreement has post-batch saving and reporting; open-ended assessment retains a reason per record but lacks a normal list return and a shipped reason-to-inference reader. See RT-RTE-adapter-training, MEM-OBJ-attention-weights, MEM-OBJ-reference-judgment and MEM-OBJ-open-judgment.

Reconciliation supersedes the epistemic Entity, reference-judgment and open-judgment declarations by the corresponding memory declarations; those canonical destinations are used here. DataPoint and judgment containers remain valid alongside their parts; no container supersession is inferred. The remaining extension-guidance qualification does not license a wired QA-only constraint: the mapping relies on field mutation/admission, not provider compliance. Extension semantics remain indeterminate, so no synthesis finding is inferred. See MEM-OBJ-entity, MEM-OBJ-reference-judgment, MEM-OBJ-open-judgment and EPI-RTE-augment-qa.

Source-read log: none. All classifications and limits use supplied records and their annotations; no target source was reread and no records or evidence were added.