Talorys memory profile

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

Comparison rationale

This profile classifies the accepted application records, not deployment behavior. All positive findings have implementation basis wired; the attributed semantic evaluation does not upgrade any classification to observed or causally supported. Missing execution evidence prevents activation, fidelity and benefit conclusions, not inspection of wired storage and consumer routes.

The application memory inventory is established across MEM-OBJ-tal-memories, MEM-OBJ-memory-reviews, MEM-OBJ-conversation-records, MEM-OBJ-conversation-summary, MEM-OBJ-semantic-vectors, MEM-OBJ-item-vectors, RT-OBJ-pending-actions, and the source-content, approval and routine paths cited in the units. These stores are not conflated with remote inference parameters. The provider adapters' uninspected internals do not imply an additional application checkpoint or establish model-weight learning; RT-CMP-workers-embedder distinguishes numerical outputs from their opaque producer.

Reconciliation supersedes the epistemic summary object with MEM-OBJ-conversation-summary, which is used throughout this profile. Its second amendment supersedes the epistemic semantic-PASS claim with MEM-CLM-eval-scope; that reported threshold result supplies no stronger basis for the memory axes. Other containers and parts remain distinct. No unresolved cross-report conflict requires choosing a stronger finding.

Natural-language data and task instructions have different consumers. Stored routine wording supplies instruction at the scheduled model, whereas selected fact/note/summary text supplies knowledge under the shipped data policy; RT-BAP-retained-context. The raw experience consumed by the summarizer or remember update has learning authority at that update consumer, separately from knowledge authority of the resulting artifact. No model compliance, improved capacity or exercised self-improvement follows. Dense vector payloads provide ranking inputs, not downstream natural-language instructions.

Consolidation describes the implemented reduction in RT-RTE-context-window and MEM-RTE-memory-write. It does not claim guaranteed semantic preservation. The indeterminate truth relation retained by the epistemic account does not require calling every compression synthesis. Historical status supports invalidation only for the bounded current-recall consumer: MEM-RTE-requested-search preserves counterpaths that still return past text. Permanent source-fact deletion supports decay as bounded forgetting: owner review replace/discard actions remove an older or newer fact, and source-memory deletion removes its primary row with associated cleanup. MEM-RTE-memory-write, EPI-RTE-lifecycle-change, and MEM-RTE-index-maintenance establish that transformation independently of historical invalidation. Remaining copies in conversations or backups limit erasure, not the source-store forgetting classification. Index reconstruction alone establishes no curation operation.

Automatic versus manual admission follows decisions, not authorship. In particular, owner selection of a backup and code admission of its rows are separate mechanisms in RT-RTE-owner-changes. Exact owner approval of a note rewrite is manual even though a model wrote the proposed body; RT-RTE-action-approval. Neither decision proves factual checking. Unknown pre-import lineage remains local to that branch rather than suppressing known authored, imported, compiled and trace-extracted paths.

The trace-learning union has a wired positive from automatic summaries and remembered facts. Summary inputs are conversational logs with recorded action-outcome feedback, not a fully faithful tool-output replay. This is supported by the memory annotations on RT-RTE-context-window and RT-RTE-action-approval, alongside MEM-OBJ-conversation-records. Other saved prose and imported prehistory lack sufficient input derivation facts for complete trace coverage. No local or system-wide no is invented in the absence of bounded absence records.

Push classifications distinguish actual unsolicited selection from requested return. Conversation and alarm identity matches select retained context, while an ID supplied to read_note is part of pull. Keyword and numerical selectors, identity-linked additions, and coarse eligibility/budget filters retain separate findings. Clef's relationship judgment changes disposition after candidate selection; it is not automatically inferred-judgment read-back selection. See MEM-RTE-memory-recall, MEM-RTE-item-recall, MEM-RTE-requested-search, and MEM-RTE-memory-write.

Source-read log: no frozen-source files were read for this classification. All warrants come from the supplied accepted records and annotations. No new records, source evidence or quotations are introduced.