CrewAI memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
All positive findings are implementation-wired classifications, not observations of operation or benefit. Axis coverage is partial because the accepted inventory names unresolved providers, application payloads and additional adapter/cache consumers. Local known units classify only their named objects or mechanisms; they do not establish a complete system-wide value list. The inventory and limits come from MEM-OBJ-memory-content, MEM-OBJ-knowledge-chunks, MEM-OBJ-flow-state, MEM-OBJ-runtime-state and their route records.
Text, dense access representations and symbolic controls are separate operative parts. Dense embeddings count as parametric form without claiming weight learning. JSON is an encoding, not a substrate: Flow/replay storage is SQLite, default checkpoints are files, and knowledge client persistence remains unresolved. A provider API does not by itself establish service-object persistence. See MEM-OBJ-memory-vector, MEM-RTE-output-replay and MEM-RTE-checkpoint-restore.
Request fulfillment and automatic model-context supply are different operations in a read-back chain. Runtime recall is pull at the recall service; task preparation is push at the later model. Model-requested Search memory remains pull. Role matching really selects retained training suggestions, whereas an explicitly requested checkpoint location alone establishes no identifier-based push. Replay has a separate downstream task-context selection after requested restoration. See MEM-RTE-memory-recall, RT-RTE-crew-training, MEM-RTE-output-replay and RT-RTE-crew-kickoff.
Learning authority at an update consumer is separated from knowledge or instruction authority at the output consumer. Retained old text can shape consolidation; recorded training feedback shapes suggestions; the latter become imperative directions. Summary history shapes a replacement artifact that later supplies continuation context. These pathways satisfy the trace-learning write-route definition without demonstrating improved capacity, truthful extraction or provider-parameter updates. See EPI-RTE-consolidation-check, EPI-RTE-plan-application, RT-RTE-crew-training and RT-RTE-context-summary.
Curation follows implemented transformations. Recency weighting supports decay; record update supports evolve; physical deletion without retained history does not establish invalidate. Extraction/consolidation labels do not prove synthesis of a new claim. The semantic relation of generalized Flow lessons remains unresolved without instances. See EPI-RTE-memory-selection, MEM-RTE-memory-maintenance, MEM-RTE-memory-encoding and MEM-RTE-feedback-lessons.
Privacy filtering supplies a bounded enforced veto on delivery, not authenticated tenancy or an invariant through update. Consolidation reconstruction can lose source/private values. Feedback pre-review may therefore consume public non-lessons as well as retained lessons; source is not an exclusive lineage selector. See MEM-OBJ-memory-metadata, MEM-CLM-retention-limits and MEM-RTE-feedback-lessons.
Canonical resolution follows the supplied reconciliation: the superseded epistemic memory-content declaration resolves to MEM-OBJ-memory-content, and the superseded epistemic feedback route resolves to MEM-RTE-feedback-lessons. Neither amendment upgrades evidence or removes opaque alternatives. The epistemic consolidation, selection and mutation stages remain valid distinct parts rather than superseded duplicates. Their narrower warrants are retained locally.
No source reads were needed or performed. All classifications use the supplied accepted records and their passages; no new records, source evidence, quotations, execution results or complete-value assertion are introduced.