WeKnora reconciliation

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

Reconciliation

Boundary and evidence basis

The four reports concern the same application-inclusive boundary and frozen revision in boundary. Its Source register supplies SRC-1 implementation and doctrine/design, with no observed run or causal experiment. Connections below use the handed records and their retained evidence; no additional source reads were made. No correction answers or declined objections were supplied to this attempt. This reconciliation therefore connects the reports as handed, without inferring a correction history.

Duplicate objects

Amendment: EPI-OBJ-knowledge-chunks is superseded by MEM-OBJ-chunk. Both identify the retained document chunks produced through parsing/indexing, with natural-language content and symbolic retrieval/lineage controls. Both ground that identity in SRC-1, internal/application/service/knowledge_process.go; the memory record additionally covers subsequent revisions and distinguishes parents from indexed child/flat chunks. The epistemic inventory and knowledge-admission ledger use the same chunk referent, not a separately persisted epistemic object. This supersession connects those findings without changing their source-warrant or semantic-fidelity limits. The broader account of revisions adds fields, not a different identity for the current chunk.

Amendment: EPI-OBJ-memory-item is superseded by MEM-OBJ-memory-item. Both identify types.MemoryItem repository rows, their natural-language Content/Topic and their separate symbolic origin, status and source links. Their common evidence is SRC-1, internal/application/service/memory/service.go and internal/application/service/memory/extract.go; the memory record adds schema/repository evidence. The epistemic record's retained status/source-ID passage and the memory record's runtime-only Inferred passage concern the same admission result. The epistemic extraction, write, confirmation and recall findings consequently concern the memory record's durable note, not a second claim store. Neither record treats source linkage as a retained argument proving the note.

Amendment: EPI-OBJ-wiki-page is superseded by MEM-OBJ-wiki-page. Both identify types.WikiPage content and provenance fields persisted by the page service with revision snapshots. Common identity evidence is SRC-1, internal/application/service/wiki_ingest_batch.go and internal/application/service/wiki_page.go. The epistemic distinction between document-level SourceRefs and ChunkRefs applies to the same page whose revision save the memory record witnesses. The epistemic wiki ledger and claim comparison retain their findings about this page; no proposition-level warrant or broader rollback guarantee follows from the identity.

These amendments leave both IDs declared and preserve each report's findings. No further exact identity is established below.

Memory routes: overlap without replacement

MEM-RTE-personal-write spans explicit directives, delayed extraction, manager operations, recurrence promotion and withdrawal. EPI-RTE-memory-extraction, EPI-RTE-memory-write and EPI-RTE-memory-review describe different functions within that account: proposal, service admission/persistence, and subsequent confirmation/rejection. Their common SRC-1 extraction/service evidence supports overlap, not identity of the whole routes. Supersession is ruled out: the broad memory route remains a valid container, and the epistemic routes retain their distinct targets and consumers.

The similarly named MEM-RTE-memory-review is not the epistemic confirmation/rejection route. It traces timed/forced expiry, demotion, model consolidation and embedding maintenance in SRC-1, internal/application/service/memory/consolidate.go. The epistemic review record instead traces confirm/reject endpoints and status/tombstone effects in the memory service. No supersession joins these two reviews.

EPI-RTE-memory-recall validly groups prompt and search consumption. MEM-RTE-personal-recall and MEM-RTE-memory-search distinguish automatic turn-time supply from foreground-requested tool delivery, with different budgets and selection inputs. The combined epistemic route acknowledges both directions. Its findings do not become wrong when separated, so it is not superseded by a split.

The reports independently converge on conditional, rather than universal, confirmation. The service admits ordinary extracted statements as active while model-flagged inferences take the pending path. The epistemic extraction prompt's event-organising example adds a permitted conjecture class; the memory account's user-only transcript selector narrows the input, not that permitted inference. These findings do not conflict. Neither API access nor the model-supplied flag establishes human review of every retained statement or evidence-certified acceptance.

MEM-RTE-retrieval-conditioning, MEM-OBJ-doc-affinity, MEM-OBJ-memory-topic and MEM-BAP-retrieval-feedback add trace-conditioned rewriting and executable ranking. They do not duplicate ordinary note recall or RT-BAP-prompt-instructions: arithmetic changes ordering at a code consumer, and query rewriting uses a separate model role. The epistemic report's admitted limits on topic/consolidation/vector coverage are uncertainty, not a denial of these positive mechanisms.

Context, document and conversation connections

The memory annotations extend RT-RTE-context-rewrite, rather than declare a duplicate route. They connect history-only checkpoint projection, raw-archive fallback and actual later reload in SRC-1, internal/agent/compaction/compactor.go and internal/application/service/agent_history.go. EPI-OBJ-context-checkpoint is its retained output object, not the transformation route, so no supersession applies. Its summary label does not establish that every eligible checkpoint is model-generated. Runtime and memory both retain the raw fallback; transient previews and the full current-turn compaction output must not be equated with the durable checkpoint. The epistemic ledger itself allows summary/raw fallback and does not assert that every checkpoint preserves meaning.

RT-RTE-knowledge-admission produces retrieval-eligible state. MEM-RTE-chunk-edit revises that output later; MEM-RTE-knowledge-read consumes it through the foreground tool. RT-RTE-rag-answer instead follows the code-selected QA pipeline. Different triggers, progression and outputs rule out route identity. The memory annotation preserves runtime's enabled-versus-enrichment distinction while adding index inputs and later divergence between saved text and index readiness. The epistemic availability-versus-truth distinction is consistent with both.

MEM-OBJ-chat-passage and MEM-RTE-chat-index concern an optional derived Q&A index with a message-search consumer. EPI-OBJ-chat-response concerns model-produced content before or apart from that indexing, and RT-RTE-agent-turn concerns the enclosing turn. Their identities are not interchangeable. No supersession equates stored raw messages, indexed Q&A, extracted personal notes or checkpoints merely because all originate around conversations.

Wiki and evaluation connections

MEM-RTE-wiki-update covers curation, direct edits, revision save and source-withdrawal work. EPI-RTE-wiki-rewrite isolates generation/retraction; EPI-RTE-wiki-guard isolates structural checking and write disposition. Their shared batch/page-service witnesses establish connected functions, not identical whole routes. The memory route remains a valid container; no split supersession is required. MEM-RTE-wiki-read is the later budgeted consumer, not another rewrite.

EPI-RTE-wiki-maintenance adds lint/AutoFix functions in SRC-1, internal/application/service/wiki_lint.go. Removing stale metadata without rewriting mixed-source prose is distinct from the queued model retraction in the memory update route. The memory report's retained retract-failure passage likewise permits old prose to remain. These are converging limits on withdrawal, not opposing deletion findings. The epistemic restriction of revert to selected content fields qualifies EPI-CLM-automatic-wiki; the memory report never promises historical provenance or placement rollback. No cross-report disagreement is established by the public claim's broader wording.

EPI-RTE-dataset-metrics explicitly identifies itself as part of RT-RTE-dataset-evaluation. Its mapping/denominator evidence adds a reference-relative precision limit where runtime left formulas uninspected. EPI-OBJ-metric-result is the produced measurement object. No identity joins the computation part, its output and the whole evaluation lifecycle. Memory's evaluation annotation agrees on in-process lookup rather than established durable retention or automatic successor adoption.

Workflow and integration dispositions

The workflow account and its annotation reuse RT-RTE-skill-install and RT-BAP-skill-image; there are no workflow declarations to supersede. Workflow adds the complete/error-free engine return prerequisite, verifier-directed repair with nil prior history, restart limits, and the pointer-publication boundary. These refine the same remote installation account. Runtime and memory already distinguish durable sandbox effects from diagnostic text and distinguish pre-switch failure from post-switch bookkeeping failure. Epistemic setup acceptance remains bounded by prerequisites rather than all functional promises. The four accounts converge without transferring ordinary chat-history assumptions to the direct installer caller.

All eight memory integration issues have compatible dispositions: supplied facts are retained; containment is not replacement; Recall's no-model comment is limited by query embedding, and capacity-only forgetting is limited by expiry/consolidation; human confirmation is conditional; withdrawal and index consistency have separate horizons; named included alternatives remain incompletely inspected; and read-back is not observed benefit. These source-comment qualifications are not corrections to a conflicting runtime finding. No supplied correction answer overturns a dependency.

No unresolved cross-report conflict is identified in the handed findings. The public claims about citations, wiki rollback and evaluation remain the epistemic report's claim-versus-route comparisons, not defects automatically attributed to another analyst. The shared lack of observed runs prevents conclusions about exercised benefit, compliance or reliable recovery; it does not erase the implemented routes. Available-but-uninspected alternatives prevent exhaustive negatives. These limits remain beside the respective records rather than being collapsed into a system-wide status.