Cognee memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
All ten axes are partial. Positive witnesses classify particular retained parts and consumers, not every shipped memory path. Cache factory alternatives are afforded rather than treated as audited storage contracts. Typed caller origins, all recall lanes, custom extraction/indexing, concurrent answer preparation, direct writes and derivative withdrawal remain unresolved in the accepted records. No absence record licenses a system-wide negative. The partial scope does not exclude an opaque alternative just because an inspected branch is classifiable.
Dense retrieval vectors and truth geometry are parametric representations, while feedback scalars, counters and epochs are symbolic localized state. Neither is evidence that embedding-provider or language-model weights change. Their graph/vector stores are substrates, not forms. See MEM-OBJ-vector-index, MEM-OBJ-truth-state and MEM-OBJ-feedback-weight.
Authority belongs to actual consumers. Retained session/run evidence has learning authority at guidance/procedure/ranking updates, while the resulting text has knowledge or instruction authority and numerical state has ranking authority downstream. Scope/tool and epoch guards establish local enforced selection, not truth guarantees or handler isolation. Conflict and supersession metadata have unresolved later consumers. See MEM-RTE-ranking-maintenance, RT-RTE-skill-revision, MEM-OBJ-skill-metadata, MEM-RTE-truth-maintenance and EPI-RTE-temporal-supersession.
Software admission is automatic on the inspected internal routes; explicit workflow start is not human approval of each retained fact. Initial supply, editing and caller-selected skill installation remain not determinable for human versus software control. Trace-fed proposal drafting is wired, but installation is a separate unresolved mechanism. Its unresolved qualifying-write assessment must not become an unconditional automatic deployed-skill update. Other session-guidance and ranking routes independently establish qualifying trace-fed updates; none establishes improved capacity. See MEM-RTE-typed-admission, MEM-RTE-session-guidance, MEM-RTE-trace-promotion and RT-RTE-skill-revision.
Consolidation describes compression or extraction of retained content; synthesis would require a new claim absent from the inputs. Curator/writer semantic indeterminacy is therefore retained rather than classified as synthesis. Exact normalized-content merging supports dedup independently of the writer's novelty request. Temporal supersession retains history and marks current reliance withdrawn by recency policy, without establishing downstream compliance. See MEM-RTE-summary-readback, MEM-OBJ-session-context, EPI-RTE-lesson-check and EPI-RTE-temporal-supersession.
Recall and load_skill are pull operations. Inside answer/update preparation, automatic selection supplies retained material to a different consumer and supports push. The writer's resolution of member_entry_ids is an identity-selected push; the model's requested skill-name lookup is not. Skills remain truncated at both handler and loop boundaries. See RT-RTE-recall-dispatch, WF-RTE-skill-loading, MEM-RTE-session-guidance and EPI-RTE-lesson-check.
The decorator's push signal is only partially classified. Recent-feedback availability, last_n and delivery truncation support coarse filtering. Query/top_k and GRAPH_SUMMARY_COMPLETION establish graph-memory selection but not its route-specific embedding operation. The unsupported inferred-embedding finding is withdrawn; the unit remains partial rather than known. Embedding findings for the separately recorded global-summary and skill-catalog selectors are unchanged. See RT-RTE-decorated-agent, including its memory annotation, MEM-RTE-summary-readback and MEM-OBJ-vector-index.
The reconciliation supersedes EPI-OBJ-learning-document with MEM-OBJ-session-learning and EPI-OBJ-truth-coordinates with MEM-OBJ-truth-state. This profile uses those canonical objects for storage, form, lineage and later consumer findings. The separate epistemic route functions remain valid; no other record is superseded. Progressive-load inspection qualifies delivery completeness without replacing the runtime authority-path identity.
No source files were read for this classification. All warrants are the supplied accepted records and their annotations, resolved through the supplied reconciliation. Evidence is implementation or doctrine/design, not observed operation or causal benefit. Missing execution evidence does not erase implemented capacity; it prevents claims of activated guidance, successful persistence, improved accuracy and beneficial learning.