Omni-Memory reconciliation

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

Reconciliation

Evidence boundary

The four supplied analyst members share the identity fixed by boundary. Its Source register contains implementation, doctrine/design and reported operation, but no observed-run or causal-experiment layer. Implemented routes, actual host consumption and demonstrated benefit remain separate throughout these reports. This reconciliation reassesses the supplied records; it uses no additional source reads. No refusal or analyst correction answers were supplied. The previous reconciliation is not evidence for the connections below.

Duplicate identities

Amendment: WF-RTE-window-handoff is superseded by MEM-RTE-transcript-window. Both refer to the same internal path: JSONL reader and saved line offset → flattened text → optional model prefix → persisted next offset. Their SRC-1 omni_memory/cli.py and omni_memory/session_memory.py passages identify the 140000-character reader threshold, 30000-character model slice and offsets[tpath] = total persistence. Both are explicitly parts of RT-RTE-capture-admission, with wired implementation and uninspected operation. Workflow's fuller account of heuristic handling, citation scanning and successful cursor persistence supplies conditions on the same omission mechanism, not a separate referent. The affected finding is capture completeness: the model may never examine text the saved cursor passes. The end-to-end capture container remains valid, both IDs remain declared, and this identity changes neither status nor evidence of operation.

Amendment: EPI-OBJ-artifacts is superseded by MEM-OBJ-generated-documents. Both identify the same .omni-memory/api-map.md and .omni-memory/linkup.md pair. Their shared generation evidence is RT-RTE-build-artifacts, including SRC-1 omni_memory/artifacts.py and its length-based replacement check. Memory additionally supplies the dashboard name allowlist in omni_memory/serve.py, covering saves and serving. These findings concern the same files, not different operative objects. The epistemic inventory and build ledger continue to assess their content relation as indeterminate; identity does not supply semantic warrant or an automatic future-memory consumer. Both IDs remain declared.

Object distinctions retained

EPI-OBJ-memory and MEM-OBJ-memory-text overlap but are not exact replacement identities. Although the epistemic record is labelled retained text, its metadata part EPI-OBJ-evidence explicitly explains that the parent denotes text with its row. Memory isolates the prose from MEM-OBJ-memory-state, which includes lineage, branch, author, protection, supersedes and quarantine/access fields. The epistemic evidence object covers a narrower symbolic part: tiers, blob maps, observations, citations and staleness. Their SRC-1 store/staleness passages support common fields, not equal scope. The row container remains valid alongside these parts; no combined finding is shown wrong by their separation, so no split supersession is made.

EPI-OBJ-candidates is transient proposed text and structured fields, not retained SQLite content. MEM-OBJ-observation-state identifies separate ledger/witness tables; RT-OBJ-inject-cache instead holds a delivery digest and prompt counter. Neither is identical to row metadata. MEM-OBJ-event-log, MEM-OBJ-savings and MEM-OBJ-code-graph retain activity summaries, estimated-cost telemetry and source-derived topology, respectively. Their substrates and consumers distinguish them from claim text and from each other. Runtime model components identify external completion backends, not any of these inputs or outputs.

MEM-OBJ-agents-context, MEM-OBJ-memory-digest and the generated-document pair are different file surfaces. Managed standing instructions, grouped memory digest and model-generated explanations have different production, selectors and consumption paths. Common Markdown encoding does not establish identity. No other operative-object supersession is warranted by these comparisons.

Route relationships

EPI-RTE-candidate-filter and EPI-RTE-filter-admission are distinct check and disposition parts of runtime capture, not aliases of each other or of the complete route. Their SRC-1 cleanup/session-memory passages agree with memory's capture annotation: software filters proposed items before writing. The transcript-window part has a different role, controlling what input reaches extraction. These parts can coexist within the valid capture container.

MEM-RTE-anchor-maintenance covers refresh/topology, trust transitions and duplicate/conflict handling. The epistemic member separates EPI-RTE-staleness-check, EPI-RTE-staleness-flag, EPI-RTE-integrity-check, EPI-RTE-tier-graduation, EPI-RTE-tier-withdrawal, EPI-RTE-merge-reconcile and EPI-RTE-merge-disposition. These narrower functions do not replace the grouped route; memory already distinguishes their tests and consequences. No split is required to repair a combined finding. They also differ from RT-RTE-use-check: capture/commit-relative anchors and persisted tier/flag changes are not retrieval-time witness pins and diagnostics.

MEM-RTE-memory-withdrawal is not identical to either EPI-RTE-eviction-check or EPI-RTE-quarantine-disposition. Memory combines reversible eligibility maintenance with store-opening migration; epistemic separates evaluation from disposition and includes explicit forget. The common SRC-1 eviction/store evidence, including restore's noise_exempt=1, establishes overlap, not equal scope. None replaces RT-RTE-local-edits, the broader caller editing route.

EPI-RTE-query-ranking is a selection part of RT-RTE-context-delivery, not an entire delivery route. MEM-RTE-recall-search returns fuller recall/dashboard payloads under different selectors and limits. Those routes and the standing-file projection are not interchangeable merely because they consume the same retained rows.

Integration dispositions and convergence

The memory checkpoint issue and workflow handoff finding qualify runtime capture without contradicting it. Runtime already records the 30000-character model limit and absence of a drain guarantee. The narrower records establish the additional omission case when cursor persistence advances beyond the prefix supplied to a successful model extraction. This does not assert loss for every batch, full-batch heuristic extraction or separate foreground JSON capture. No observed loss incident is supplied.

Memory's selection issue agrees with epistemic query ranking: empty-query/no-context browsing may sort by recency, rather than applying universal trust weighting. Memory's initial-eight limit is not a final count after global/sibling append. Runtime delivery already distinguishes those append paths and the first-row budget exception. The supplied records therefore support bounded selection, not chain completeness or guaranteed transmission of stored reasons.

The authority issue connects three distinct claims rather than merging them: RT-CLM-verified-context concerns enforced host use; EPI-CLM-earned-verified concerns tier qualification; EPI-CLM-hygiene-truth concerns retention protection. Runtime's manual-source --verified, dashboard fields and trusted import are separate paths from automatic citation-plus-fresh-hash promotion. Memory maintenance and epistemic graduation agree on that predicate and its limits. Initial inferred AI/doc rows do not meet its stated-tier condition. Source-label authority, file identity and citation are not thereby semantic truth or successful task outcomes.

The import issue refines RT-RTE-shared-import: selected fields survive import, but observation/use/rollback state is not fully restored. Existing-ID skip prevents replacing previously imported content, and disappearance from a later shard does not revoke a local copy. This is compatible with retained provenance and automatic clamping, not a denial of them. Local withdrawal does not establish distributed revocation.

The projection-consistency issue is also a qualification, not a disagreement. Runtime local edits describe row updates and later projections without requiring every file to refresh simultaneously. Memory specifies that local remember/build do not immediately rewrite managed AGENTS.md; workflow specifies that capture has no single transaction across rows, feedback, documents and cursor. Different projections omit different caveats. Neither row archival nor regenerated files withdraw already returned host-window copies.

The savings issue concerns MEM-OBJ-savings: assumed baseline costs and recording before auto-output suppression prevent interpreting the ledger as measured delivered blocks or causal savings. No other supplied record depends on that interpretation. Remaining parser, symbol-propagation, map, browser and retained-flow questions stay as local inspection limits. They neither establish missing operative parts that must be invented nor justify system-complete absence claims.

Remaining disagreements

No cross-report disagreement was identified that requires an Unresolved conflict: paragraph. No supplied correction answers establish a corrected or declined objection to adjudicate. The exact duplicate identities are stated above; other overlaps retain distinct scopes, parts or valid containers. The reports' shared uncertainty about deployment is not a disagreement. RT-BAP-agent-context remains a delivery/interpretation path with advisory host force, not evidence of compliance, improved capacity or exercised self-improvement. This reconciliation changes no analyst finding and makes no verification or acceptance judgment.