Agentic Local Brain memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The inventory follows retained parts and their later consumers, not the whole application's capability surface. Every finding is architectural wiring or explicitly labelled guidance, never observed deployment or causal benefit. The units preserve unknown remote/provider state and untraced alternatives rather than extending one inspected route to every caller.
Dense document vectors remain parametric representations even when serialized as JSON or packed float bytes in SQLite. Localized scores, IDs and membership fields are symbolic. Chunk vector storage uses the vector substrate; SQLite dense-vector rows do not become a separate vector database merely because they encode vectors. See MEM-OBJ-chunks, MEM-OBJ-document-embeddings and MEM-OBJ-topic-clusters. Provider calling records do not identify provider-side accumulated memory, despite naming model components and service roles.
The imported document body, compiled representations and authored model text have different derivation stages. Generic caller identity cannot establish human authorship or manual admission. Generated missing metadata, mined products and conversation turns have automatic software admission; archive adoption separately requires an operator's replacement decision. See RT-RTE-file-admission, RT-RTE-item-update, RT-RTE-backup-restore and MEM-RTE-note-admission.
Requested knowledge/recommendation/article returns are pull. Inside an answer request, loading session-matched prior turns and automatically supplying recent-view influence are separate contextual push operations. No unsolicited scheduler is required for this distinction. The prior-turn push uses an actual session match and then coarse turn/budget filtering. Recent-view enrichment uses coarse event selection; its conditional ID-based score boost is not independently established as identity-selected delivery of chunks. These mappings use RT-RTE-chat-query, MEM-OBJ-rag-conversations and MEM-RTE-history-selection, without turning every automatic retrieval selector into push.
The conversation chain supports a wired trace-learning witness: recorded turns enter generation, assistant prose is automatically retained on applicable paths, and later invocations receive selected prior text. The original trace source is conversational session content. This does not assert a continuation-summary mechanism, model training or improved capacity. Merely storing raw reading events does not close the same derived-artifact chain: the recommendation profile is transient. Consequently those units remain unresolved rather than supplying a system-wide negative. See RT-RTE-chat-query, MEM-OBJ-rag-conversations, MEM-BAP-retained-context and MEM-RTE-history-selection.
Consolidation has a deterministic sentence-selection witness and a separate model-guidance finding. Wiki wording requesting synthesis does not establish a new claim absent from inputs: generated semantics remain indeterminate. Recompilation and centroid updates establish revision of retained products; title cleanup establishes a bounded duplicate criterion. Reading-event time decay changes influence, not event deletion. See EPI-RTE-extractive-summary, RT-RTE-wiki-compile, EPI-RTE-numeric-selection, EPI-RTE-title-cleanup and MEM-RTE-history-selection. No synthesis finding is inferred from an unknown model output; deletion without demonstrated historical retention is not classified as invalidation.
Reconciliation supersedes the entity-object declaration in epistemic with MEM-OBJ-entities; the profile therefore cites that canonical object. The valid article, answer/confidence and HTTP-withdrawal/title-cleanup containers retain distinct parts and are not collapsed. The budget disagreement between RT-RTE-chat-query and MEM-CLM-budget-scope prevents a universal budget guarantee, not the recorded history/retrieval existence findings. This profile retains the hierarchical/flat distinction and makes no whole-message cap claim.
All ten axes remain partial because their named inventories contain unresolved alternatives, origins or consumers. Positive findings are retained with their local evidence layers; no strongest witness upgrades unknown alternatives. Entity-card lineage, complete collector and maintenance paths, exact user-note/source-dictionary consumption, archive payload inventory, provider accumulated state and some topic/graph semantics prevent stronger assessments. No bounded absence record establishes a trace-learning no across those units.
Source-read log: none. Classification used only the supplied records and reconciliation; no source facts or quotes were added and no report verification was reopened.