Atomic memory profile

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

Comparison rationale

All findings classify the accepted records, not new source evidence. Reconciliation makes no supersessions: wiki articles remain a valid container, while pending proposals and archived versions keep their distinct consumer and admission units. Their knowledge-at-reader paths do not grant automatic live-model authority. The current-run CitationLedger and ToolRegistry are excluded from accumulated memory; selected persisted citations are included with their products.

Database retention does not imply vector encoding is symbolic or vectors are provider weights. Dense chunk embeddings map to parametric form and vector substrate. PostgreSQL atom support gives a relational substrate witness but does not settle the substrate inventory of every other retained part. Chunk, tag, conversation, wiki live/candidate/archive, report and vector storage units retain their inspected positive findings with partial coverage: their included alternative backend branches are not completely classified by the accepted records. This is a substrate-coverage limit, independently of transaction or cleanup parity. No SQLite-only complete classification of those object-wide inventories is asserted.

Requested corpus reads are pull. History supplied automatically to a model producing a new turn, report source seeding and host-selected prompts/taxonomy are push operations within a requested enclosing workflow. This operation-level distinction follows the profile type; it does not turn the report's descriptive phrase about no independent unsolicited history push into a pull-only classification. Wiki research chooses retained chunks for another synthesis consumer; identity, model judgment and subsequent coarse budget filtering are separate selectors. Query-driven keyword/vector retrieval remains pull and does not alone establish lexical/embedding push.

Automatic model-derived admissions are distinct from unresolved generic caller control. The saved-answer helper does not prove human per-save consent or exclude it. No manual finding is inferred from a user-facing surface or workflow initiation. Similarly, archive reads do not establish human checkpoint replacement. Static default prompts are not accumulated memory, while changed retained prompt values supply standing instruction authority. The retained custom tagging prompt separately supplies instructions to the tagging provider, not only to chat; its path is classified below alongside the chat-prefix path.

Tag merging supports category deduplication/consolidation; existing atom/wiki revision supports evolve. Archived prior wiki prose loses current live reliance while remaining historical, supporting invalidate on that replacement path. No command name or request alone establishes decay, promote or semantic synthesis. The accepted epistemic generation records leave wiki/report output semantics indeterminate; neither controlled synthesize nor consolidation is therefore established for those generation units.

The report route records a final writer consuming research messages and tool results, automatic durable finding admission and later product readers. That local chain supports trace learning and learning authority at the writer independently of knowledge authority at downstream readers. Its session/tool origins coexist with nontrace source snippets and unresolved original corpus provenance. It neither establishes improved capacity nor attributes model-weight learning. Raw chat retention, context truncation and diagnostic evaluator metrics are not independent qualifying derived updates merely because they record activity. Saved-answer and remaining alternative admission paths retain their uncertainty rather than becoming bounded negatives.

Every axis remains partial because the accepted record inventory explicitly leaves included client, adapter, strategy and maintenance alternatives unresolved. Positive existence warrants are not complete-value assertions. No supplied evidenced-absence record warrants a system-wide negative. Wiring is not observed activation or causal benefit. The generated-product reader's numeric deletion can also remove assertion content; product knowledge authority therefore describes delivered context, not faithful semantics or transitive source truth.

The coarse mixed representational-form units retain supported existence findings, not complete part-level consumer equivalence. Conversation role/content reaches provider replay while attached tool/citation metadata has separate storage readers; wiki candidate prose reaches review while base IDs/timestamps reach freshness checks. Withhold part-level and consumption-path comparisons that assume every field in either mixed unit reaches one consumer or receives one interpretation.

Supporting distinctions: MEM-RTE-tag-compaction, MEM-RTE-answer-save, MEM-OBJ-wiki-proposals, MEM-OBJ-wiki-versions, RT-RTE-report-finding, EPI-OBJ-generated-prose, EPI-RTE-generated-delivery and EPI-RTE-retrieval-scoring.

Source-read log: none. All classification warrants come from the supplied accepted members and their annotations; no source ambiguity required a new read. No analyst records, quotations or evidence were added.