Engraph reconciliation

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

Reconciliation

Boundary and evidence

The members share the service boundary and frozen source in boundary and its Source register. Engraph is the shipped vault service, not the external conversational agent. Implementation and doctrine/design establish source-level routes and declared expectations; no supplied execution or comparison evidence establishes operation or benefit. That shared limit is not a disagreement about implemented capacity. This reconciliation reads only the supplied members and previous reconciliation, adds no source inspection and declares no analyst records. No correction answers or declined objections were supplied; the current records, not the previous reconciliation's interpretation, govern the connections below.

Duplicate identities

Amendment: EPI-OBJ-identity-facts is superseded by MEM-OBJ-identity-fact. Both identify the SQLite identity_facts rows with tier/key/value/source/time fields, populated by extraction and subsequently formatted. The epistemic INSERT passage and memory source-less upsert passage identify that retained object (SRC-1, src/store.rs, src/identity.rs). The epistemic record also discusses separately configured L0 identity, but distinguishes it from the rows; this amendment does not identify L0 config with database facts. Truth-apt category assessments, source loss and extraction-time versus represented-state recency remain findings on the shared referent, not replacement values.

Amendment: EPI-OBJ-health-report is superseded by MEM-OBJ-health-report. Both identify the current-call HealthReport, not a saved evaluation. The epistemic struct passage and memory stale-function passage establish the common diagnostic output and unevaluated stale slot (SRC-1, src/health.rs). Retained unresolved-link input is an input to that report, not an alternative identity. The epistemic structural-proposition assessment and memory diagnostic-freshness limits both survive.

Amendment: EPI-RTE-identity-extract is superseded by MEM-RTE-identity-context. Both trace full indexing/setup or explicit identity refresh through tier-1 clearing, heuristic category extraction, SQLite retention and later identity formatting. Their matching selectors, note-count horizons, omitted source fields and lack of ordinary watcher re-extraction identify the same progression (SRC-1, src/identity.rs, src/indexer.rs, src/watcher.rs). The epistemic OOO counterexample and checkbox-stripping passage assess that progression rather than establish another route. The amendment preserves extraction, retention and later consumption as separate consequential functions. Both declarations retain wired implementation and uninspected operation; retained candidates do not thereby become accepted claims.

Amendment: EPI-RTE-placement-correction is superseded by MEM-RTE-placement-feedback. Both trace a detected move, provenance/suggestion comparison, centroid addition/subtraction, correction logging, suggestion-field stripping and later semantic placement. Their recognized-label constant, watcher adjustment passage and distinction between centroids and correction logs establish identity (SRC-1, src/placement.rs, src/watcher.rs, src/store.rs). The epistemic overlap with runtime admission and refresh does not distinguish it from the memory subroute of refresh. Both declarations retain wired implementation and uninspected operation. The epistemic member's more detailed indexing/rebuild limits do not replace the memory finding or imply a global feedback reset.

Containers, parts and non-identities

EPI-OBJ-vault-content and MEM-OBJ-vault-notes overlap at disk Markdown but are not interchangeable as declared. The epistemic object also covers symbolic and numerical representations and person/project associations; the memory object isolates source files. That valid container can coexist with MEM-OBJ-file-record and MEM-OBJ-chunk-record, which distinguish access metadata, lossy snippets and vectors. No supersession of the container by those parts is asserted. These objects are inputs or results of RT-RTE-note-admission and RT-RTE-index-refresh, not duplicate routes.

The identified placement-feedback route does not replace all of RT-RTE-index-refresh, whose other indexing paths need not process qualifying moves, or RT-RTE-note-admission, which also covers initial placement and edits. MEM-OBJ-folder-centroids is durable numerical selector state; MEM-OBJ-placement-corrections is a raw correction log. Their different later consumers rule out identity with each other, the feedback route, or loaded model weights.

EPI-RTE-content-withdrawal and MEM-RTE-note-withdrawal share archive/delete/cleanup/clear branches but are not exact replacements. The memory progression includes unarchive and restored content; the epistemic record treats unarchive as a separate recovery request while assessing withdrawal and missing-path cleanup. Their overlapping destructive effects do not establish identity of the complete progressions. The memory declaration is a lifecycle part of RT-RTE-note-admission; the epistemic declaration spans admission and index cleanup. Neither replaces those runtime containers. Their limits on soft delete, archive and auxiliary cleanup are compatible, not conflicting complete-erasure claims.

EPI-RTE-orchestration-parse and EPI-RTE-confidence-normalize are parts of RT-RTE-query-search, not duplicate complete searches. EPI-RTE-health-report consumes indexed state but performs diagnostics rather than index refresh. MEM-RTE-vault-readback covers explicit read, section, person/project, list and overview alternatives, not another ranked search or RT-RTE-topic-context. Identity extraction spans setup and refresh without replacing either whole runtime route.

The remaining memory objects—profile and CLI telemetry, alongside the distinguished file/chunk/centroid/log parts—have no duplicate declarations in the supplied members. Existing cache, topic context, setup and migration referents are annotated rather than re-declared. The epistemic claims about vault knowledge, confidence, diagnostics and placement learning do not duplicate runtime's read-only or write-consistency claims. The workflow declares no records and adds fields to supplied identities. No further supersession is asserted.

Disagreement assessment and integration dispositions

No unresolved cross-report conflict is established by the current supplied records. In particular, the previous reconciliation's centroid conflict no longer describes the pinned epistemic finding. EPI-RTE-placement-correction now limits indexing replacement to represented first-component groups from indexed_rel_paths, states that rebuild does not clear every centroid row, and retains the possibility that nested parent-folder feedback rows survive. MEM-OBJ-folder-centroids and MEM-RTE-placement-feedback likewise restrict replacement to touched rows and reject a coherent full-vault aggregate. Their first-component/chunk-count versus parent-folder/file-mean witnesses agree (SRC-1, src/indexer.rs, src/watcher.rs, src/store.rs). The old disagreement is therefore not carried forward. Neither report warrants global feedback invalidation or coherent centroid recovery, and both preserve the positive update-to-later-placement pathway.

The memory integration issues have these remaining dispositions:

  • Identity freshness: MEM-RTE-identity-context and EPI-RTE-identity-extract agree that ordinary watcher branches do not regenerate tier-1 facts. RT-RTE-index-refresh says retained index state feeds identity extraction, not that every event performs it. No dependent every-event guarantee remains to conflict with that qualification.
  • Lifecycle split: the two withdrawal records retain differing source and derivative effects for archive, soft delete, hard delete and restoration. Unresolved auxiliary cleanup limits complete erasure; they do not contradict the bounded runtime admission account. No undeclared part is required to sustain a complete-erasure claim, because none is made.
  • Feedback eligibility: the feedback records agree on recognized provenance labels, suggestion metadata, semantic-placement bypasses and separate log/centroid consumers. Runtime deliberately leaves feedback revision to memory rather than claiming all moves are human corrections. The centroid-unit mismatch is an agreed limit, not a disagreement with a runtime aggregate invariant.
  • Health: MEM-OBJ-health-report, EPI-RTE-health-report and EPI-CLM-health-diagnostics converge on the stale detector's stub. The epistemic record additionally distinguishes absent advertised CLI dispatch from positive MCP/HTTP reporting. This is a public-claim versus implementation mismatch, not an analyst conflict or an automatic freshness approval gate.
  • Untraced parts: MEM-OBJ-cli-event establishes storage/pruning capability without a traced learning consumer. Profile detection, graph formatting, auxiliary cleanup and external-client handling remain local inspection limits. No supplied dependent finding requires those unknowns to be positive or treats them as evidenced absence.
  • Duplicate handling: the amendments above identify the shared identity and feedback referents; the non-identities preserve distinct objects, alternatives and containers. Memory's intention not to duplicate supplied runtime records is compatible with duplicates between the independently declared memory and epistemic records.

Independent convergence and dependency limits

The workflow account independently traces create within RT-RTE-note-admission. Its commit-before-rename ordering, attempted cleanup and persistent effects before response delivery agree with RT-CLM-write-consistency. Workflow delivery is serialized metadata, not semantic acceptance, automatic read-back or the user's conversational answer. MEM-RTE-vault-readback supplies a later requested consumer without becoming a mandatory completion gate.

Memory's annotation on RT-RTE-topic-context agrees with runtime's basic-search selector, disk reread, byte-based trimming and non-strict cap. Epistemic assessment adds possible qualification loss, not a claim that every truncation changes truth. Memory's annotation on RT-OBJ-llm-cache agrees on query-hash selection without model/age invalidation. EPI-RTE-orchestration-parse distinguishes generated-text parsing from cache deserialization, so syntax admission is not incorrectly inherited as a semantic check on every cache hit.

Fixed inference parameters in the runtime components and durable move-fed centroid adaptation concern different numerical state and do not conflict. EPI-CLM-placement-learning is assessed against that wired adaptation pathway, without promoting persistence to measured improvement. Across members, operational write permission, retained content, retrieval ranking, advisory delivery and epistemic acceptance remain distinct. None supplies observed operation or causal comparison that upgrades these routes to successful recovery, client compliance, improved answers or beneficial learning.