semquery memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
This profile classifies accepted records, not a new source inspection. Reconciliation establishes no supersessions; the original declaring-member citations remain canonical. No source files were read for this profile, and no evidence or records were added.
The operative corpus is split into text, lexical terms, dense vectors and symbolic access/maintenance state. SQLite and vector storage coexist: a vector access structure is not a different claim about the physical database. Dense vectors are distributed-parametric representations, not model weights; the retained chunk text separately supplies reference content and reranker input. See MEM-OBJ-chunks, MEM-OBJ-fts-chunks, MEM-OBJ-vec-chunks and MEM-OBJ-document-paths.
Automatic selection within a requested search or ask is pull, not a second push operation. Thus lexical and embedding relevance signals are not asserted on the push-only signal axis. Concrete configuration/cache loads also fulfill component requests; unknown custom delivery mechanisms remain separate rather than borrowing their coverage. The recorded selector and context budget do not establish complete model visibility: GGUF can discard leading formatted tokens. See RT-RTE-hybrid-search, RT-RTE-config-admission and RT-RTE-answer-generation.
The use-produced metadata has routing force at maintenance consumers. Content hashes and baseline specifications select processing rather than checking the truth of document claims. Similarly, citation filtering acts on transient answer metadata, not on a retained memory admission criterion; it adds no validation or enforcement finding for the memory inventory. See MEM-OBJ-document-paths and EPI-RTE-marker-check.
Deduplication is implemented through shared text identity. Withdrawal supplies deletion-based forgetting, not history-preserving invalidation. These classifications retain the limits of shared ownership, stale empty-content transitions and differing cleanup predicates; they do not claim reliable mirroring or successful deletion in a run. Forced reembedding and rebuilds alone do not establish semantic consolidation or synthesis. See RT-RTE-index-admission, MEM-CLM-chunk-lineage and MEM-CLM-empty-retention.
External saved configuration replacements can provide standing instructions, but unchanged shipped guidance is not accumulated memory. The wired consumer supports instruction authority conditionally on such a replacement, not observed editing or compliance. Unknown human involvement cannot be relabelled manual merely because the CLI has an operator. The automatic document admission mechanism is independently known. See MEM-OBJ-config, MEM-OBJ-system-prompt and RT-RTE-index-admission.
No accepted evidenced-absence declaration covers the learning inventory. Reports' bounded statements that answers are not written back and logs are not used for automatic policy revision therefore do not license a controlled trace-learning "no". Nor does raw diagnostic retention establish "yes". Opaque injected implementations and external revisions remain unresolved, and trace source cannot be marked globally inapplicable. Exact model identity, tokenizer contents and initial authorship also remain local gaps, not reasons to discard known transformation findings. See MEM-OBJ-semq-log, RT-RTE-answer-generation, RT-RTE-config-admission, RT-CMP-fast-embed, RT-CMP-fast-reranker and RT-CMP-gguf-llm.
All positive findings describe wiring. No observed consumer compliance, retrieval quality, improved capacity or causal benefit is inferred. Known form or substrate does not authenticate artifact provenance, and known automatic writes do not resolve every external admission decision.