RightMemory memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The comparison covers retained content and its access/maintenance state, not every storage object in the runtime. File persistence, repository history and graph organization coexist; dense retained vectors are classified separately from external inference weights. Natural-language content, fixed-rule identity/control fields and distributed numerical vectors have different consumers. Provider-managed continuation is included but unresolved rather than inferred from readable snapshots. See MEM-OBJ-memory, MEM-OBJ-retrieve-context and MEM-OBJ-vectors.
Automatic Update is a qualifying trace-learning route when its submitted evidence records agent work and user redirection: a semantic model admits durable guidance/cases and later retrieval supplies them to a host. That supports a wired local "yes", not observed improvement, universal trace provenance, weight training or successful self-improvement. Ordinary document import and raw history retention are different paths. Opaque continuation processing and mixed submissions remain separate unresolved units. See MEM-OBJ-update-evidence, MEM-OBJ-agent-corrections, RT-RTE-isolated-update and RT-BAP-host-context.
Curation uses conservative evidence layers. Models are afforded keep/merge/revise and pruning operations through the executable roles and their semantic instructions; the source does not supply compliant output traces. Extractive distribution selection is deterministic and wired. Generative view production and Insight requests do not establish a newly synthesized claim without a recorded input/output relation. A faithfulness defect would not itself prove synthesis. See RT-RTE-maintenance-cycles, EPI-RTE-view-build and MEM-OBJ-insight-logs.
Requested retrieval remains pull even when models or numerical ranking select its answer. Internal snapshot/diff/pending/changed-delivery additions are a distinct automatic supply operation. The identifier finding concerns saved delivery coverage identifying changed previously delivered sources, not merely IDs in a requested selector. Opening-context projection fulfills an explicit consumer request; graph adjacency does not independently establish identifier-based push. See RT-RTE-agent-retrieval, MEM-OBJ-retrieve-context and MEM-RTE-opening-context.
Write agency follows admission: native/internal semantic edits and refresh are automatic despite upstream workflow initiation. Human editor decisions and explicit inbox-proposal approval establish local manual routes. Generic approval callers and initial feedback population remain unresolved. Permission flags enforce sharing/answer eligibility; they are not factual acceptance. See RT-RTE-native-assistant, RT-RTE-pursuit-editor, MEM-RTE-guidance-review, MEM-RTE-view-refresh and EPI-OBJ-view-permission.
Reconciliation supersedes the epistemic correction-case, edit-feedback, inbox, Insight-log and shared-view container declarations with the corresponding memory declarations. It supersedes the epistemic provider-question route with the memory question-view route. All classifications use these canonical records; the declared epistemic view-document, retriever-policy and permission parts remain distinct. The broader epistemic Memory container is not substituted for any single memory part. These identity resolutions change citation targets, not evidence strength or coverage.
Every axis retains unresolved included coverage rather than inferring complete enumeration from a positive witness. Actual deployments, initial content control, provider payload internals, full remote transport/withdrawal behavior and additional feedback adoption are not resolved by these records. No bounded absence record supports a trace-learning negative. No observed or causal basis is asserted.
Source-read log: no source files were read. This classification uses only the supplied accepted members and reconciliation; no additional quotations, records or evidence were introduced.