sqlite-graph-memory reconciliation

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

Reconciliation

The supplied reports share the frozen boundary and its Source register. This reconciliation connects their current findings without adjudicating support or changing their statuses. No analyst correction answers or declined objections were supplied. Implementation, attributed doctrine, host activation and causal benefit remain separate evidence questions.

Duplicate supersessions

Amendment: EPI-OBJ-turns is superseded by MEM-OBJ-turns. Both identify the SQLite turns rows produced by the hook and read by the viewer, not separate stores. Identity evidence is their common SRC-1 anchors, src/sqlite_graph_memory/turnstate_hook.py and src/sqlite_graph_memory/turnstate_show.py, the same ask/summary/decision and session/tool/file/command/evidence fields, and passages showing concatenated summary production and viewer selection. The epistemic inventory, ledger and bounded conclusion therefore assess the memory record's retained object. The epistemic deductions about decision-looking text and attempted file operations remain findings about that object; supersession does not promote those events to verified decisions or successful effects.

Amendment: EPI-RTE-diagnostic-checks is superseded by MEM-RTE-vault-health. Both describe caller-triggered sgm-stats/sgm-doctor, graph/date/environment checks, threshold verdicts, stdout/JSON and exit status, without automatic repair or persistent report history. Common SRC-1 evidence is src/sqlite_graph_memory/vault_stats.py and src/sqlite_graph_memory/doctor.py; the records retain passages showing report emission and exit-code consequences. The epistemic ledger's separate checking and disposition functions describe this same route family. Its domain limits remain applicable: graph counts, stale-date classifications and cache/index presence do not validate note truth or successful recall. Neither route establishes exercised watchdog use.

Overlap without interchangeable referents

EPI-OBJ-note-snippets spans acquired note bodies and stored excerpts, describing metadata and separate numerical storage alongside them. MEM-OBJ-notes identifies source files and live links; MEM-OBJ-brain-meta identifies the persisted pickle's snippets and access/display fields. The scopes overlap but neither memory record alone replaces the epistemic referent. No identity supersession is made. The acquisition-to-excerpt account remains a valid broader description: the common indexer passages support frontmatter removal and clipping, and both reports preserve the possibility that qualifications do not reach the host.

EPI-OBJ-recall-state covers persistent vectors together with invocation-local candidates, scores and formatted output state. MEM-OBJ-brain-e5 identifies only the persisted matrix. The broader object and its numerical material can coexist; no supersession is made. Neither object is identical to the inference components RT-CMP-e5-model or RT-CMP-rerank-model. MEM-OBJ-brain-manifest instead identifies reuse/regeneration metadata consumed by the indexer. Weights, output vectors, query state and freshness metadata remain distinct.

EPI-OBJ-evaluation-records identifies gold, answers and metrics. It is not interchangeable with the setup-time whole-note map MEM-OBJ-benchmark-raw, or with MEM-OBJ-ab-recall, which retains within-query list-comparison telemetry. Related inputs and outputs do not establish identity. EPI-OBJ-diagnostic-report is diagnostic output, not the diagnostic route superseded above. Object/producer overlap requires no additional supersession.

Connected findings and integration dispositions

The workflow annotates RT-RTE-mcp-recall and RT-CLM-scoped-return, without declaring another route. Its rejection and recovery account shows that header-only output can pass the parent's nonempty-file check. Runtime's successful-return guarantee concerns the designated output file rather than stdout; it does not guarantee a positive hit count. The epistemic comparison likewise distinguishes transport success from warrant. These findings converge without an inter-report conflict: a scoped return is not semantic acceptance, and its empty-file error is not a complete zero-hit detector. Host recovery remains outside the package after handled failure.

The current RT-RTE-turnstate-capture distinguishes the installed console entry point from module/script invocation. Its pyproject.toml and example-registration passages identify direct main invocation; its wrapper passage limits exception suppression to running as main. Thus storage failure can propagate on the installed path after offset advancement, while the module/script wrapper suppresses it. Memory's MEM-OBJ-turns, MEM-OBJ-turnstate-offset and capture annotation require neither universal suppression nor reliable persistence. Epistemic's EPI-OBJ-turns concerns parsed content and viewer losses, not an exception-handling guarantee. No dependent finding in those reports requires the installed hook to suppress errors. The current accounts remain compatible; external host response to propagated failure is unresolved, not evidence of successful recovery.

Memory's annotation on RT-RTE-recall-options adds PPR target-resolution differences and cache limits. Its SRC-1 passages from src/sqlite_graph_memory/ppr.py and src/sqlite_graph_memory/index_notes.py distinguish stem lookup and optional hash fingerprinting from one-hop resolution. Runtime already distinguishes candidate expansion from ordering and process-local cache horizons. This qualifies any claim of one shared graph, without disputing the runtime account. Different entity policies in CLI, A/B, ablation and benchmark remain explicit in RT-RTE-evaluation-paths. EPI-CLM-measurement-scope compares those differences with public measurement claims; it neither contradicts runtime nor treats every metric as invalid.

EPI-RTE-extract-answer, EPI-RTE-answer-abstain and EPI-RTE-bench-score are declared parts of RT-RTE-evaluation-paths. Extraction, answer admission and reference checking have distinct functions; the runtime container remains valid and is not superseded by them. EPI-RTE-bench-gate adds disposition of result claims for reporting, not another name for evaluation execution or RT-RTE-change-admission. Memory's full-body benchmark channel and epistemic's extraction account agree that benchmark answering cannot be treated as production snippet delivery. Reference-dependent metric derivation and result consistency checks do not establish observed gains, factual truth or automatic policy adoption.

RT-BAP-context-delivery and RT-BAP-temporal-instructions have different consumers and channels: returned snippets versus operator-installed prescriptions for an external writer/reader. EPI-CLM-search-layer concerns note recall's functional scope, not the transport guarantee. RT-CLM-temporal-scope compares prescribed temporal reliance with implemented replacement access. Memory and epistemic independently retain clipping, frontmatter omission and continued access to closed content. This doctrine/implementation discrepancy is shared across reports, not a disagreement between them.

Memory's remaining integration issues retain bounded dispositions. Original note provenance and per-content human control remain unresolved for MEM-OBJ-notes and RT-RTE-index-admission; starting an indexer settles neither. Refresh scheduling and actual host compliance are external. The distinct database defaults for telemetry, capture and viewing qualify a single-file deployment assumption without contradicting table identity. Transcript evidence pointers depend on surviving external files, and viewer clipping does not convey full provenance. The offset is a separate control object, not another ledger. None of these uncertainties establishes automatic later model adoption or conflicts with the current capture record. Evaluation-result conditions, maintainer decisions and deployed successor use remain uninspected, even though evaluator and release wiring are described.

No unresolved inter-report disagreement or required undeclared part was identified in the supplied records. Faithfully scoped uncertainty remains with the affected records rather than becoming a conflict or a stronger status. No frozen-source read was needed for this reconciliation.