Smriti MCP reconciliation

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

Reconciliation

The reports share the frozen shipped-artifact boundary and its Source register. Their implementation accounts concern SRC-1; none supplies observed target execution or causal benefit. The external caller owns semantic retention decisions, consolidated prose and eventual host prompt use. Unresolved caller identity and host consumption limit all four accounts without constituting disagreements. No correction answers or refusal were supplied to this attempt. This reconciliation uses the pinned members without additional source reads.

Duplicate supersessions

Amendment: EPI-OBJ-memory-body is superseded by MEM-OBJ-memory-body. Both identify MemoryDocument.body as caller-supplied natural-language content serialized in local Markdown, with external interpretation after retrieval. Their identity evidence is SRC-1, src/smriti_mcp/frontmatter.py and src/smriti_mcp/store.py: the epistemic record quotes the body field and replacement assignment; the memory record describes the same retained field and quotes full-note reload. The epistemic inventory and routes citing the epistemic ID therefore concern the memory record's body, not a second payload. Unknown factual warrant, generic authorship and actual external consumption remain unchanged.

Amendment: EPI-OBJ-memory-meta is superseded by MEM-OBJ-memory-meta. Both identify MemoryDocument.meta, the YAML mapping and its derived runtime representations, with displayed prose assertions and fixed-rule metadata consumers. Both retain the same confidence-score passage from SRC-1, src/smriti_mcp/store.py; the epistemic record additionally cites src/smriti_mcp/server.py, and the memory record identifies serialization in src/smriti_mcp/frontmatter.py. The epistemic inventory and authority ledger refer to the same operative metadata, not independently warranted confidence or verification results. This supersession does not identify the metadata with either generated index: those have separate construction and consumption boundaries.

Amendment: EPI-OBJ-trace-event is superseded by MEM-OBJ-trace-events. Both identify the record_trace event payload persisted as NDJSON, including supplied observations, remembered-content copies and generated consolidation reports. SRC-1, src/smriti_mcp/store.py supplies the epistemic record's JSON append passage and the memory record's grouping passage. Singular versus plural labels do not establish different referents. The epistemic trace-review and retention findings concern this same raw-event population; neither record authenticates origin or factual support, and the memory record preserves the runtime's trace-path qualification.

Amendment: EPI-RTE-access-reinforcement is superseded by MEM-RTE-access-reinforcement. Both isolate recall marking or direct mark_accessed, durable timestamp/count/salience updates and later affinity ranking. Both declare implementation wired and operation uninspected, and both cite the identical capped salience assignment in SRC-1, src/smriti_mcp/store.py; the memory record also quotes timestamp/count persistence and cites src/smriti_mcp/server.py. The epistemic behavior/policy-adaptation ledger row concerns this same numerical feedback route, not a second adaptation mechanism. The memory record retains the direct trigger, pre-mark recall summaries and later search/recall/remember ranking consumers. Identity supplies no additional evidence of changed selection, correctness or improved capacity.

These supersessions connect duplicate identities only. Both IDs remain declared; no analyst finding or status is replaced by reconciliation. The epistemic records' statements of no supplied object counterpart compare against runtime objects, not the separately pinned memory declarations. The memory report's no-duplicate integration statement likewise compares its new records against supplied runtime findings. Neither statement rules out these cross-specialist identities.

Containers, parts and distinct records

RT-CMP-memory-store is the executor, not any retained payload. RT-OBJ-server-instructions is static shipped guidance, distinct from the retained body, metadata and trace objects above. RT-BAP-tool-guidance describes its host-facing consumer channel and force; it is not a duplicate of the guidance object or subsequent recalled content.

RT-RTE-memory-write remains a valid container alongside its declared parts: EPI-RTE-shape-check, EPI-RTE-write-admission and EPI-RTE-lifecycle-disposition. They respectively isolate structural checking, branch admission and caller-directed withdrawal/replacement. Their shared source, SRC-1, src/smriti_mcp/store.py and src/smriti_mcp/server.py, does not make those functions identical. No split supersession of the container is needed. EPI-RTE-content-retention spans note writes and trace consolidation and expressly is not wholly part of either; it is not a duplicate of admission, checking or either larger route.

RT-RTE-context-recall remains a valid read/write container. EPI-RTE-recall-selection isolates selection/delivery; the superseded access-reinforcement pair isolates durable feedback and also admits a direct marking trigger. Selection and feedback differ in effects and horizon, so they are not identities. The memory annotation supplies response omissions, graph-selection inputs and approximate budget details without redeclaring the parent.

RT-RTE-trace-consolidation remains a valid production/review/admission container, with EPI-RTE-trace-review isolating advisory grouping and recommendations. Raw events are its inputs, not a second route. The workflow annotation supplies the caller-led handoff, independent invocations and recovery consequences rather than a duplicate declaration.

RT-RTE-store-maintenance remains a valid container alongside EPI-RTE-maintenance-check and EPI-RTE-maintenance-rewrite. Diagnostics and rewriting have distinct results and force. The rewrite record expressly identifies direct create's description fallback as a counterpath to a claim, not another maintenance invocation. MEM-OBJ-markdown-index and MEM-OBJ-machine-index are distinct derived artifacts: saved Markdown has explicit load delivery, while transient machine dictionaries feed fresh retrieval and saved YAML has no inspected reload consumer. Neither is identical to a producing route or the note metadata it derives from. No missing part is identified by the pinned reports that requires superseding these valid containers.

The three epistemic claims—EPI-CLM-consolidation-provenance, EPI-CLM-summary-omission and EPI-CLM-recall-bundle—have different claimed operations. None duplicates RT-CLM-file-controls, which concerns ordinary file controls and their counterpaths rather than provenance, description lineage or task-context selection. The workflow declares no new records requiring identity amendments.

Independent convergence and integration dispositions

Consolidation provenance and semantic authority. Runtime's RT-RTE-trace-consolidation, the memory annotation, EPI-RTE-content-retention and the workflow account converge on caller-authored final prose, unverified trace-reference strings and branch-specific provenance: create/update apply prepared metadata, append does not. Their load-bearing passages concern SRC-1, src/smriti_mcp/store.py and src/smriti_mcp/server.py. This disposes of the broad wording in EPI-CLM-consolidation-provenance as a recorded doctrine-versus-route qualification, not a conflict between analysts. Review is requested, not enforced; persistence is not evidence-consuming acceptance.

Omitted descriptions. The memory integration issue and EPI-CLM-summary-omission independently identify README/input-description wording contradicted by the remember create branch's delegation to create. Their passages cite SRC-1, README.md, src/smriti_mcp/server.py and src/smriti_mcp/store.py. RT-RTE-memory-write and the runtime account already allow description derivation during creation, so no dependent runtime finding requires the omission claim to be true. EPI-RTE-maintenance-rewrite retains a source-level truncation counterexample and an indeterminate semantic-preservation assessment; the memory report's deterministic-prefix lineage statement does not assert universal semantic preservation. The source doctrine discrepancy remains local to the claim and metadata lineage, not an unresolved cross-report conflict.

Recall selection and later consumption. RT-RTE-context-recall, EPI-RTE-recall-selection, the memory read-back account and workflow recall annotation agree that requested delivery ends at the caller. SRC-1, src/smriti_mcp/store.py supplies their budget and snippet passages. Full get remains available; compact recall omits complete bodies and provenance. The memory integration issue about related metadata does not dispute runtime traversal: both identify body wikilinks/backlinks as graph inputs, rather than a second traversal from metadata associations. Approximate budgeting and access-driven future ranking qualify EPI-CLM-recall-bundle without establishing observed benefit or a truth oracle.

Withdrawal, durability and recovery. EPI-RTE-lifecycle-disposition, the memory write annotation and RT-RTE-memory-write agree that archive/supersession preserve historical notes and default recall exclusion is not universal revocation. RT-CLM-file-controls and the workflow recovery account agree on per-note replacement without composite rollback, using SRC-1, src/smriti_mcp/store.py. A failed event append can follow a persisted note; returned exceptions do not establish undone effects. The workflow's caller-inspection recovery consequence does not strengthen the runtime guarantee.

Maintenance and feedback limits. The memory maintenance annotation, EPI-RTE-maintenance-check, EPI-RTE-maintenance-rewrite and RT-RTE-store-maintenance agree on invoked diagnostics or deterministic changes, with hidden-path and dry-run limits in SRC-1, src/smriti_mcp/store.py and src/smriti_mcp/fix_frontmatter.py. Diagnostic derivation concerns code-relative structural predicates, not note truth. Automatic access reinforcement is separately wired; neither that feedback nor caller-led trace review establishes improved capacity, autonomous semantic criticism or exercised theory revision. These are independent property limits, not competing system-wide statuses.

No unresolved conflict between the pinned analyst findings is identified. The documented claim/implementation mismatches above remain explicit, as do unknown caller provenance, external consumption and missing execution/comparison evidence. No declined cross-report objection was supplied; none is treated as an established defect.