xMemory memory profile

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

Comparison rationale

The profile classifies the accepted records, not an execution. All positive findings are wired: the retained members supply implementation witnesses but no observed run or causal experiment. Partial assessments preserve named alternative mechanisms and provenance gaps rather than downgrading those positive witnesses.

The memory inventory separates prose, symbolic access fields and dense vectors even when they share a file. MEM-OBJ-retrieval-vectors supports parametric representation, not a model-weights storage finding or a parameter-training claim. MEM-OBJ-memory-graph supports a separate diagnostic graph/file/RAM substrate; the adaptive reader consumes theme and kNN files instead. Generic injected repositories and provider state remain unresolved rather than inheriting default classifications from readable results.

The reconciliation supersedes the epistemic originals, semantic-content and theme-summary duplicates with MEM-OBJ-original-messages, MEM-OBJ-semantic-content and MEM-OBJ-theme-summary. Its owner-deletion supersession resolves to MEM-RTE-memory-withdrawal. This profile cites those canonical declarations. The broader memory episode/title, topology and vector records remain valid alongside narrower epistemic parts; overlap is not a second independent witness.

Retained text has different authority at different consumers. It supplies knowledge to answering and scoring through RT-BAP-memory-reader. Retained experience also feeds durable updates through RT-RTE-semantic-admission and RT-RTE-hierarchy-admission, supporting learning authority at those update consumers. Neither this authority nor the positive trace-learning classification asserts improved capacity. Similarity comparators have a different role: EPI-RTE-duplicate-check and EPI-RTE-semantic-disposition establish validation plus an insertion veto, not truth acceptance. Persisted confidence without a consuming rule supplies no additional authority.

Requested QA remains pull throughout ranking, expansion, hydration and return. Automatic supply of retained inputs to later update-model invocations is a different operation: episode-event continuation supplies an identified episode, prediction receives selected prior statements, and theme summarization receives selected member strings. Those are push units within update chains, not a claim of unsolicited foreground-agent delivery. The annotations on RT-RTE-semantic-admission and RT-RTE-hierarchy-admission, together with MEM-OBJ-theme-summary, establish embedding, fallback availability/order and identity selectors. Query-time NLL judgment in RT-RTE-adaptive-reader is not counted as a push signal merely because code delivers requested results.

Consolidation is classified from the implemented condensation/extraction routes, not a claim of semantic equivalence. EPI-RTE-comparison-extraction leaves entailment versus ampliation unresolved, so it cannot establish synthesis. Theme regeneration supports local evolution even though it can omit old member text. Explicit deletion supports local forgetting, not history-preserving invalidation or complete disappearance from all consumers; MEM-RTE-memory-withdrawal retains the residual-file/cache and concurrent-writer limits.

The positive trace-learning route is conversation-to-retained-summary/statement-to-later-consumer, not raw log storage or offline benchmark feedback. Its original conversational turns support session-logs. Captions and search text have insufficient recorded origin for tool-traces or other additional source values. RT-RTE-offline-evaluation and EPI-RTE-lexical-evaluation describe post-answer reports without an automatic memory-revision consumer; those artifacts are excluded from this accumulated-memory comparison boundary. EPI-OBJ-predicted-episode and EPI-OBJ-entropy-score are transient computation, not independently retained learning artifacts.

Source-read log: none. Classification used only the supplied accepted members and reconciliation; no source ambiguity required a direct source read.