xMemory reconciliation

Type: agentic-system-analyses/types/agentic-system-reconciliation-report.md

Reconciliation

The four reports share the boundary: the shipped construction, maintenance, retrieval, answer and evaluation loop, including its drivers. The Source register supplies implementation and doctrine/design, plus attributed README operation claims, but no observed-run or causal-experiment evidence. This reconciliation uses the supplied members only; no source reads were needed. No analyst correction answers or declined objections were supplied to this attempt.

Duplicate identities

Amendment: EPI-OBJ-original-messages is superseded by MEM-OBJ-original-messages. Both identify Episode.original_messages: conversation prose with role/time envelopes, retained in episodes and used for comparison and conditional answer hydration. The memory record retains the conversion witness; the epistemic record supplies the unknown external-warrant assessment. This is one trace object, not two independent evidence sources. The epistemic inventory and ledger's import/hydration findings concern that same object; supersession changes neither its provenance limits nor consumer authority.

Amendment: EPI-OBJ-semantic-memory is superseded by MEM-OBJ-semantic-content. Both explicitly identify SemanticMemory.content, rather than the entire stored envelope. Both trace its prediction/comparison or extraction lineage and its subsequent prediction, theme and answer consumers. Their conversion evidence agrees that source episode IDs accompany statements and confidence is assigned as a constant. The epistemic inventory, extraction, disposition and retention findings concern this same content. The supersession does not turn operational insertion into truth acceptance or absorb metadata into the prose part.

Amendment: EPI-OBJ-theme-summary is superseded by MEM-OBJ-theme-summary. Both identify ThemeNode.summary, its generated natural-language text, theme-file substrate and adaptive-scoring consumer, distinct from the final answer's inputs. Their cited summary-assignment and refresh evidence concerns the same field. The epistemic qualification about topic labels versus recurring-fact assertions and the memory qualification about current-batch refresh coexist. The affected hierarchy-transformation and fidelity findings do not gain either semantic validation or complete member coverage from this identity.

Amendment: EPI-RTE-owner-deletion is superseded by MEM-RTE-memory-withdrawal. Both routes identify the explicit MemorySystem.delete_user_data progression: owner selection, buffer/store deletion, index clearing and component flags. Both retain the load-cache pass witness and distinguish administrative removal from all-consumer withdrawal. The memory record additionally discusses adjacent individual-delete and cleanup capabilities; those qualifications do not create another owner-deletion route. The epistemic recovery-ledger finding remains administrative removal, with no truth evaluator and no universal revocation guarantee. Both implementation statuses remain wired and operation statuses remain uninspected.

Overlap without supersession

EPI-OBJ-episode-content names Episode.content; MEM-OBJ-episode-content groups title and content. Exact whole-record identity is ruled out by that scope difference. The former isolates content within the latter's valid prose container. Their shared content lineage is not independent confirmation of semantic fidelity. No split supersession is needed: the supplied findings do not make the title/content grouping wrong. Neither object duplicates its producer, RT-RTE-conversation-admission.

EPI-OBJ-selection-state isolates ThemeNode.semantic_ids; MEM-OBJ-theme-topology also includes theme identity, count and neighbor structure. This is membership-part overlap, not identity of the full referents. Likewise EPI-OBJ-embedding-vectors inventories episode/statement/theme embeddings, while MEM-OBJ-retrieval-vectors expressly includes centroids as well as search and summary embeddings. Their overlapping dense representations share numerical selection uses, but the records do not establish coextensive scopes. No whole-record supersession is asserted. Neither is identical to the producing model, RT-CMP-embedding-client, or to symbolic MEM-OBJ-semantic-knn.

The explicitly declared parts EPI-RTE-episode-normalization, EPI-RTE-comparison-extraction, EPI-RTE-duplicate-check, EPI-RTE-semantic-disposition and EPI-RTE-semantic-retention are not identities of their runtime containers. Normalization isolates a post-generation transformation within conversation admission. The four semantic parts separate generation, checking, insertion disposition and persistence within RT-RTE-semantic-admission. Their distinct targets and consequences explain the epistemic ledger's separation of functions. The runtime containers remain valid; no combined finding requires their supersession by parts.

EPI-RTE-lexical-evaluation is a declared part of RT-RTE-offline-evaluation, not a replacement for the parent. It adds directly inspected lexical-domain and aggregation details. EPI-OBJ-qa-result isolates the generated response; EPI-OBJ-reference-answer isolates the imported reference. Their shared result-file substrate does not make them identical. Nor do transient EPI-OBJ-predicted-episode and EPI-OBJ-entropy-score duplicate retained episode text or retrieval vectors.

MEM-OBJ-memory-metadata, MEM-OBJ-memory-graph and MEM-OBJ-original-index have no same-referent declaration in the other reports. Metadata accompanying prose, a diagnostic graph export, and callable original-message selectors are not alternative names for the runtime routes. EPI-CLM-faithful-hierarchy concerns intact units and faithful organization; RT-CLM-retrieval-purpose concerns uncertainty-guided expansion and benefit. A shared README paragraph does not make these claims identical. Workflow declares no competing records and annotates existing routes instead.

Integration dispositions and convergence

Theme coverage and export alignment. Memory's first integration issue qualifies RT-RTE-hierarchy-admission through MEM-OBJ-theme-summary: fresh manager caches and a new-batch text dictionary can leave old members outside regenerated summaries. Runtime asserts summary generation and structural maintenance, not complete text coverage. The epistemic fidelity claim remains a design claim, not an independently validated result. These positions can coexist. MEM-OBJ-memory-graph separately distinguishes retained diagnostic nodes/edges from live selector state; it does not contradict runtime's statement that the adaptive reader uses separate files.

Dispatch names and budgets. Memory's second and third issues refine RT-RTE-baseline-search and RT-RTE-adaptive-reader in its Annotations. NOR-LIFT dispatch reaches ordinary Chroma ranking in the inspected wrapper; runtime's dispatch inventory does not assert otherwise. Original-message limits are positive-only caps on episode blocks, not tokens, individual messages or a zero-disable control. Baseline Chroma results do not set the expansion flag. Runtime describes flagged hydration subject to configuration; workflow explicitly says a positive cap. No supplied finding requires an incompatible universal cap or unconditional hydration conclusion.

Retention, repair and withdrawal. Memory's fourth and fifth issues align with the epistemic semantic-part and deletion records: backend flags do not guarantee global invalidation; constant confidence and append-oriented insertion do not establish factual repair. RT-RTE-semantic-admission expressly leaves detailed revision classification to the specialists. Its use of admitted or accepted statements describes save decisions, not an evidence-consuming truth-acceptance decision. Memory's sixth issue retains unresolved embedding compatibility, injected backends and unspecified helper consumers. These are scoped limits, not disagreements or evidence that the alternatives are absent.

Adaptive selection versus benefit. Runtime's surrogate-score account, memory's adaptive annotation, epistemic's System-claim versus route comparison and workflow's Workflow account converge on optional, code-controlled context selection, not verified uncertainty reduction. Negative thresholds, fallback, summary-level expansion tests and differences between scoring and final-answer inputs qualify the README claim. Agreement among reports does not add observed operation or causal support to RT-CLM-retrieval-purpose.

Delivery versus acceptance and recovery. Workflow adds that answer exceptions may become empty saved responses, while escaping worker exceptions prevent the normal save call. These are alternate outcomes within RT-RTE-adaptive-reader, not contradictions of runtime's ordinary save progression. Its exhausted-flush branch qualifies entry to hierarchy maintenance without denying earlier episode writes. All three lenses place offline scoring after answer generation and outside an automatic memory-revision loop. The lexical evaluator's reporting consumer is compatible with the absence of an automated admission veto; an operative report is not an answer-acceptance gate.

Authority and coverage. RT-BAP-memory-reader supplies the model-interpreted text authority shared by the specialist accounts. Memory refines it by actual consumer: themes enter scoring, while selected semantics, episodes and possibly originals enter final answering. Numerical structures enforce selection rather than acting as prose instructions. The source-wiring account remains separate from successful influence, reliable truth and learning benefit. RT-ABS-action-dispatch remains bounded to inspected default routes and coexists with RT-RTE-provider-admission's injected-code and trusted-loader surface; it establishes no process sandbox.

No unresolved cross-report disagreement was identified in these pinned members. The duplicate supersessions above resolve identity, not support or status. No additional required undeclared part was established from the supplied findings. The verifier still judges the records' support; remaining coverage, execution and causal limits stay attached to the affected findings rather than becoming a system-wide negative.