Virtual Context memory profile

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

Comparison rationale

This classification uses accepted records only. No source file was read, no record was declared, and no report verification was reopened. Every positive finding is wired implementation, not observed activation or causal benefit. Every axis is partial because the retained inventory includes unresolved alternatives; a known local unit is not a complete system-value claim.

Parts and canonical identity

The reconciliation supersedes EPI-OBJ-fact with MEM-OBJ-fact. All fact-object citations use that canonical record. This affects form, substrate, lineage, authority and the conversation-derived trace chain without upgrading fact truth or coverage. Synopsis, structured envelope and tag rollup remain different operative parts. Card entry/check/integration records likewise remain parts of a wider cache, not competing identities.

Relational storage, dense vector storage and archive files classify retained objects rather than encoder weights or serialization. The checkpoint's known in-process/Redis alternatives do not classify its opaque provider-owned persistence or the curation provider. Media metadata does not establish the opaque image payload's form or lineage. See RT-OBJ-engine-state, MEM-OBJ-payload-archives and MEM-OBJ-search-vectors.

Independent effects and transformations

Conversation experience has learning authority at its automatic derivative update; delivered statements have knowledge authority at a later answer consumer. The card wrapper separately supplies standing response directions. Neither route proves compliance, semantic truth or improved capacity. The card's disabled default does not remove its wired path. See RT-RTE-memory-maintenance, EPI-RTE-card-integration and EPI-BAP-card-influence.

Fact/refusal state can both test admissibility and enforce a veto. Those consumers are not a general truth oracle. Source-responsive revision and withholding support evolve and invalidate. Extraction of standing facts from retained messages supports consolidate; a dropped negation does not establish synthesize. Cold paging reduces the retained working-set tier, not the source archive. Requested deduplication or route naming does not substitute for an inspected retained-memory transformation. See MEM-RTE-fact-admission, EPI-RTE-fact-withholding and RT-RTE-paging-loop, including its memory annotation.

The explicit human-admin edit/verdict path supports manual admission, while a generic API actor remains indeterminate. Starting import, compaction or posting does not make downstream software admission manual. Provider persistence and other repair/replacement mechanisms stay unresolved. See RT-RTE-record-edits, MEM-RTE-fact-admission and RT-RTE-import-history.

Selection and trace chains

Inbound work triggers automatic push to the answer model. Explicit paging, recall and restoration are pull. Requested API delivery is not counted as an additional push solely because software returns text. Tag/actor identity matching, inferred vector/model selection and coarse window/budget gates have distinct local selectors; no lexical finding is invented for incompletely inspected retrieval strategies. See MEM-RTE-inbound-context, MEM-RTE-payload-restore and RT-BAP-engine-watermarks.

Automatic durable summaries, facts and cards qualify as trace learning because later invocations consume them as context or guidance. Session turns retain session-log origin even when imported. Dense-vector computation establishes compiled numerical lineage, not encoder training. Raw tool/media archiving alone does not establish a qualifying learned artifact; lacking bounded absence records, it also cannot establish a local no covering all its alternatives. See MEM-OBJ-source-content, MEM-RTE-card-refresh and MEM-OBJ-search-vectors.

Engagement's automatically retained question/day history shapes future candidate, draft and posting eligibility. Its staging/claim activity supplies event-stream input locally. Arbitrary candidate and model-callable origins remain unresolved. The personal-versus-timed fidelity distinction changes draft admission but does not eliminate the independent retained-history consumer. A dry-run return alone has no runner ledger write. See RT-RTE-engagement-stage and WF-RTE-draft-selection.

Full MCP resource/search routes, repair/backfill bodies, evaluator feedback, initial provenance, opaque providers and complete card serving predicates remain limited by the accepted accounts. Known witnesses do not turn those included parts into absence, inapplicability or complete enumeration. Evaluator cache reuse and answer scoring establish neither benchmark benefit nor a live behavior-update consumer for general diagnostic scores. See MEM-RTE-evaluator-reuse, EPI-RTE-answer-scoring and RT-RTE-mcp-tools.