Siftly memory profile

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

Comparison rationale

Boundary and canonical identities

This is a revision-2 classification of accepted records, not a new inspection of Siftly. The substrate is the retained bookmark system, not all runtime storage. Process queues and shipped prompts are not independently accumulated memory. Updated settings are included only where later resolvers consume them; external context/checkpoint retention remains an unresolved alternative at the adapter boundary. The actual executors' parametric form does not establish use-accumulated parametric memory. See RT-RTE-model-dispatch, RT-RTE-settings-change and MEM-OBJ-raw-json.

Reconciliation supersedes the epistemic entity, image-tag, semantic-tag and category-description object identities with MEM-OBJ-entities, MEM-OBJ-image-tags, MEM-OBJ-semantic-tags and MEM-OBJ-category. Values above cite those canonical objects. The additional direct category writer and empty-store seeding remain qualifications of that same category object, not a new memory surface. Its incomplete admission assessment prevents complete write-agency, lineage and curation coverage. Membership-only and membership-plus-enrichedAt records are not whole-record aliases. Neither are query matches and the complete response cache. The wider memory records retain their distinct scopes.

Consumption, selection and authority

SQLite FTS5 is a lexical access structure, not vector memory, and the mindmap is a display projection, not a graph substrate. Language tokens can have natural-language force at a model and symbolic force at a fixed selector. Classification therefore separates these consumption paths rather than treating JSON as a semantic form. See MEM-OBJ-bookmark-fts and MEM-RTE-browse-memory.

Category descriptions supply standing classification instructions. Tags and source text supply context and ranking inputs. Allowed slugs supply both admissibility criteria and enforced exclusion. Processing markers route later work; they are not evidence of correctness. No blanket learning authority follows from these effects. In particular, success at categorization can set enrichedAt even after semantic enrichment failed. The accepted retry qualification changes routing coverage, not the existence of retained metadata. See RT-BAP-category-guidance, RT-BAP-enrichment-search, MEM-OBJ-bookmark-category and EPI-RTE-semantic-tags.

A requested search is pull even when automatic lexical/model selection fulfills it. A different operation automatically supplies retained metadata and category guidance into pipeline model calls. That operation is push, with bookmark identity and broad availability/state predicates kept separate. Neither classification asserts an unsolicited injection into an external agent. Category/all-record availability is coarse selection; a requested CLI ID is not a push identifier signal. See RT-RTE-search-content, RT-RTE-enrich-categorize and RT-RTE-cli-export.

Transformation and uncertainty

Reduced persisted exports support consolidation; assignment/default overwrites support evolution; cache TTL/eviction supports decay. Duplicate import IDs, URL reuse and full index rebuilding alone do not establish semantic deduplication. Deletion without retained history is not the controlled invalidate operation. Actual vision/enrichment outputs are absent, so semantic novelty or faithfulness cannot be inferred from a generation route name. No synthesis finding is added from that uncertainty. See MEM-OBJ-export-copy, MEM-OBJ-search-cache, MEM-OBJ-category, EPI-RTE-image-analysis and EPI-RTE-semantic-tags.

There is no accepted evidenced-absence declaration for trace learning. The reports' feedback-search limits do not license a local no, much less a system-wide no. Imported post content is not automatically a session/tool trace, and attempted-analysis markers are not established durable experience-derived guidance. Without a qualifying local write chain or complete negative coverage, trace source cannot become inapplicable. Opaque external retention remains uninspected for both axes. This withholds a classification rather than asserting absence or requiring demonstrated improvement for a wired yes. See MEM-OBJ-raw-json, MEM-OBJ-image-tags, MEM-OBJ-bookmark-category and RT-RTE-model-dispatch.

Axis assessments preserve local findings without claiming complete system unions. Automatic/manual differ by admission mechanism and evidence layer; the human UI affordance does not upgrade generic API callers. External file reading is afforded, not observed. No finding is observed or causally supported: the supplied records establish source wiring and explicit affordances, not successful use, truthful annotations or improved performance.

Source-read log

No target-source reads were made. Only the supplied invocation, job, types, contracts and pinned members were read. No additional evidence or records were introduced.