Synto memory profile

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

Comparison rationale

This classification uses accepted records and their memory annotations. No target source was reread and no new evidence was added. Reconciliation supplies no supersession: containers and narrower parts remain canonical. The findings describe implementation or afforded later host consumption, not observed operation or causal benefit.

Objects, provenance and control

SQLite and files are distinct persistence substrates; retained relations also provide graph organization. Dense embedding arguments establish numerical form, not an inspected vector backend. The injected store remains in the inventory on every affected axis rather than disappearing behind a known file/SQLite witness. See MEM-OBJ-rag-embeddings and MEM-OBJ-relations. Provider inference weights are outside this use-accumulated memory boundary, not an asserted negative about provider learning.

Imported notes, model/code-derived knowledge, authored human criticism and captured call/activity outputs have different derivation paths. Admission is independent: a human can supply an import or reject a draft while software admits extracted segments or a feedback-conditioned rewrite. Save flags similarly enable automatic content admission. Initial external/seed provenance and injected-store control remain unresolved. See MEM-OBJ-source-documents, MEM-OBJ-rejections and MEM-RTE-synthesis-save.

Consumer effects and retrieval

Feedback has instructional authority at the writer and learning authority at the durable-revision update consumer; the resulting article supplies knowledge at later answering. These are separate effects, neither demonstrating better capacity. Identity denials can both validate a candidate and enforce its rejection. Retained catalogue text, aliases and relation confidence are ranking/routing inputs even when scoring code is fixed. Generated pack guidance has only afforded instruction/routing force at the excluded host; file generation alone does not prove host loading. See RT-BAP-rejection-guidance, MEM-OBJ-concept-entities, RT-RTE-query-answer and RT-BAP-pack-guidance.

A question requests retained page context, and source tools explicitly request passages: those chains are pull. Automatically supplied concept hints, rejection feedback and compatible prior results are separate push operations. Exact cache/checkpoint identity matching selects delivered payloads; merely returning a requested segment ID does not establish identifier-based push. No opaque RAG selector is inferred from the visible write. See MEM-RTE-source-passages, MEM-RTE-checkpoint-resume and MEM-RTE-response-reuse.

Transformations and uncertainty

Consolidation describes retained-source compression, not guaranteed faithful generation. The source-native synthesis command does not establish the controlled curation value synthesize without recorded novel-claim evidence. Merge supplies a local identity-deduplication witness; publication raises a draft tier; retirement withdraws live use while retaining bodies. No universal retraction or full-state rollback follows. See RT-RTE-concept-curation, RT-RTE-draft-publication and EPI-OBJ-query-answer.

Ordinary Q&A remains filesystem/export-accessible and title-resolvable, but refreshing the index does not enumerate that tree. Synthesis has explicit later catalogue admission subject to the newest-25 limit. This accepted qualification changes neither record identity nor the positive later-consumer routes. See EPI-RTE-answer-retention, MEM-OBJ-synthesis and MEM-OBJ-index.

The review-feedback chain supports trace learning as an automatic experience-fed durable revision, not exercised self-improvement. Its source is human criticism of a generated draft; the records do not establish that criticism as a conversational session log, recorded tool request/result, timestamped runtime event stream or linked state/action/outcome trajectory. Trace source therefore remains not-determinable instead of manufacturing an event-stream label. Cache/checkpoint reuse and metrics reports also retain their unresolved qualification and consumer limits; raw activity retention alone does not warrant a learning union. See WF-RTE-interactive-review, RT-BAP-rejection-guidance, MEM-RTE-response-reuse and MEM-OBJ-metric-events.