ATLAS memory profile

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

Comparison rationale

The comparison follows the accepted local object and route inventory, not all runtime storage. MEM-OBJ-record-payloads, MEM-OBJ-extract-log and MEM-OBJ-progress-state are distinct database parts. RT-OBJ-knowledge remains their separate filesystem counterpart; MEM-OBJ-messages is browser memory. Dynamic configuration is included only at its recorded later memory-operation routing consumers in RT-RTE-config-reload. Static schemas and chat instructions are not accumulated memory. Nonpersisted query calculations and developer diagnostic returns do not add stores or learning routes; see EPI-RTE-derived-results and EPI-RTE-evaluator-tools.

Reconciliation supersedes the duplicate epistemic contradiction object with MEM-OBJ-contradiction-block. All classification citations use that canonical object; EPI-RTE-contradiction-marking remains a distinct insertion route. The supersession changes no value. Conflict prose supplies a warning to readers, not an enforced withdrawal of either assertion. A returned applied flag can follow a missing-target no-op; neither that flag nor a model request for duplicate avoidance upgrades a transformation to observed execution. RT-CLM-additive-knowledge also prevents treating add-only policy as a universal preservation invariant.

Selection and requested return are different operations. RT-RTE-knowledge-enrichment pushes lexically selected full files into an update prompt. RT-RTE-chat-tools pushes an availability index but pulls requested contents. A literal path on that pull or in the index does not establish identifier-targeted push. MEM-RTE-conversation-replay pushes retained role/content history while omitting prior tool-result evidence. Coarse filtering remains separate from lexical targeting and from actual content delivery.

Trace learning is a write-and-later-consumer classification, not a performance verdict. RT-RTE-extraction-summary transforms retained runtime metadata into possible durable knowledge; it does not reconstruct extracted entities from counts. MEM-OBJ-progress-state also records automatically derived outcome checkpoints whose later use changes acquisition selection. Those generated control artifacts meet the trace-fed behavior-shaping boundary independently of natural-language knowledge. Raw log/history retention alone does not. The chat-save branch has a wired history/tool-feedback-to-file-to-later-reader chain in RT-RTE-chat-tools; actual saved content and mixed input provenance remain unresolved. The raw-replay account does not supply a canonical bounded-absence warrant, so it establishes neither a resolved local no nor trace-source inapplicability. The positive union retains the separately warranted yes branches. None of these findings establishes improved future capacity or provider parameter training.

All axes remain partial because local positive witnesses do not resolve unavailable external module memory or provider-private retention/context management. RT-RTE-module-lifecycle and RT-CMP-llm-client bound these alternatives without classifying them from process capabilities or numerical model computation alone. Initial deployment content, generic caller identity, actual model outputs and semantic transformations remain local uncertainties. Known store wiring does not resolve those distinct lineage, agency or transformation questions. All positive implementation mappings are wired unless explicitly limited to claimed duplicate avoidance; no supplied record supports observed or causal memory benefit.

Source-read log: none. Classification used only the supplied accepted members and reconciliation; no frozen-source ambiguity required a direct source read. No records or new evidence were added.