sift-kg memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The comparison distinguishes filesystem persistence from the operative graph and parsed process objects. It also separates symbolic graph/control fields from interpreted assertions, excerpts and guidance. The same JSON or YAML container can support both forms without making either a model-weight store. The optional embedding encoder and its temporary cluster state do not establish accumulated vector memory (RT-OBJ-graph-data, RT-CMP-sentence-transformer). Provider state remains an explicitly unresolved alternative, not an asserted second memory implementation (RT-CMP-llm-client).
The supported local lineage is compilation from documents and earlier artifacts, plus human-authored review decisions. Initial custom content remains unresolved. Software admission remains automatic even when a person starts the command; human review keys and explicit edits are separate manual mechanisms (MEM-OBJ-document-extraction, RT-RTE-review-decisions). No evidence upgrades wired routes to observed operation or causal benefit.
Authority follows actual consumption. Retained schema guidance instructs later extraction; symbolic schema fields validate or enforce particular build/review rules. Decision statuses validate executable choices and enforce graph changes. Graph degree and names influence relative selection without requiring a learned ranker. Reasons shown to a reviewer do not become engine-executed justifications (RT-BAP-domain-guidance, EPI-OBJ-domain-vocabulary, RT-OBJ-review-files, RT-RTE-agent-retrieval).
Curation includes actual duplicate merging, support aggregation, revision and deletion. Deleting a rejected graph relation is forgetting at that store, not history-preserving invalidation or synchronized withdrawal from every artifact. Rebuild can restore it from extraction caches. Generation and style correction do not establish synthesis of a new claim; semantic preservation is locally unsettled (MEM-OBJ-relation-support, RT-RTE-apply-decisions, EPI-RTE-graph-cleanup, RT-RTE-narrative-generation).
External retrieval is pull. Internal chunk-call construction automatically supplies reusable domain guidance and the first-chunk summary, so that separate consumer path is push. This does not reinterpret requested stdout as unsolicited host delivery. Its recorded selection is coarse availability/configuration-based supply, not embedding retrieval or targeted identity matching. The summary is within-job retained context consumed by later calls, not a disk checkpoint (RT-RTE-agent-retrieval, RT-RTE-schema-discovery, RT-BAP-domain-guidance, RT-RTE-document-extraction).
Trace learning remains not determinable. Document-derived reusable guidance is a behavior-shaping artifact, but its recorded source is not established as an activity trace. Output-fed prose repair likewise does not establish a recorded experience-to-durable-update chain. There are no canonical absence declarations to support a local negative, and no qualifying trace source is established. Unknown API/provider alternatives cannot be discarded to derive a system-wide negative (MEM-OBJ-document-extraction, RT-RTE-schema-discovery, RT-RTE-narrative-generation, RT-CMP-llm-client).
Reconciliation supersedes the epistemic extraction-object and graph-support identities with MEM-OBJ-document-extraction and MEM-OBJ-relation-support. Those canonical objects carry this profile's citations; their source-native identity resolution changes no evidence strength or consumer force. The valid domain and graph containers, their material parts, and the separately retained prose artifacts remain distinct. The narrative's afforded reading is not upgraded by the wired description sidecar route (MEM-OBJ-narrative, MEM-OBJ-entity-descriptions).
Source-read log: none. Classification used only the supplied accepted members and reconciliation; no source ambiguity required a new frozen-source read.