MentisDB report verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

The five submitted members support publication of a bounded implementation account, with the local inventory limit below. I read every supplied report, its type, the reconciliation and the shared record contract. All four answers files are empty; the artifact check reports none, so there are no structural failures to route and no claimed corrections to accept on trust. Independent source inspection used SRC-1 from the Source register. Checkout HEAD equals the reviewed revision and git status --short is empty. No target, example, test, benchmark or provider was executed. Source implementation and shipped doctrine do not become observed operation through this verification.

Runtime, objects and authority

I traced service append through core construction, canonical storage, index updates and deferred flushing, and reopening through integrity checking and migration. RT-OBJ-thought and RT-RTE-memory-roundtrip correctly distinguish structural admission from content validation. Storage precedes fallible derived work; errors can follow admission. Hash verification checks the stored sequence, not factual truth or authenticated authorship. Legacy migration can rewrite files. The bounded guarantee wording is supported. SRC-1, src/lib.rs, src/server.rs.

The natural-language content part in MEM-OBJ-thought-content and EPI-OBJ-thought-content is distinct from the envelope's symbolic consumer. A full read of the Python adapter confirms raw-message writes, unused session targeting, twenty-hit type-filtered retrieval without current-input text, role inference from agent-ID equality, exception-to-empty behavior and no-op clear. Consequently the reports' rejection of transcript fidelity and session isolation is warranted. Prompt-string delivery supports RT-BAP-retained-context, not observed reliance. SRC-1, pymentisdb/langchain.py, src/lib.rs.

RT-OBJ-core-skill, RT-BAP-core-skill and RT-CLM-startup-protocol preserve the difference between available instructions and append invariants. Reading the complete skill and local stdio dispatch confirms host-owned startup, checkpointing, identity discipline and local saving. Direct append has no prerequisite search check. Enabled HTTP authorization classifies explicit request chains; subsequent ancestor discovery and ranked search do not impose the same check on every discovered chain. Local protocol execution bypasses HTTP middleware. RT-RTE-access-control therefore supports conditional authorization, not universal isolation. MEM-CLM-scope-visibility concerns another mechanism: tags and supplied filters cannot establish the skill's identity-based visibility promise, especially given the Python and recent-context selectors. SRC-1, MENTISDB_SKILL.md, src/bin/mentisdb.rs, src/server.rs, src/lib.rs.

I independently read the extraction prompt, request construction, parser and transformer. RT-CMP-extraction-model, RT-RTE-memory-extraction, RT-CLM-extraction-review, EPI-OBJ-extracted-thoughts and EPI-RTE-extraction-schema have supported scopes. Model output is transient; type/nonempty-content checking and confidence clamping do not check truth. The service's separate extraction and append operations do not establish actual caller review. Indeterminate semantic generation is appropriately distinguished from deterministic trimming/defaulting. SRC-1, src/llm.rs, src/server.rs.

RT-CMP-fast-embed-provider, MEM-OBJ-embedding-vector, MEM-OBJ-retrieval-state and MEM-RTE-derived-update distinguish inference machinery, numerical output and symbolic retrieval state. Provider inference, embedding-input assembly, append synchronization and implicit-overlay updates support that lineage and selection authority. Neither the model-family enum nor embedding-space label pins an exact artifact digest; no adapter training operation is established. Freshness and rebuild descriptions remain conditional, not exhaustive assurances about metadata changes, WAL replay or all backends. SRC-1, src/search/fastembed_provider.rs, src/lib.rs, src/search/vector.rs, src/search/implicit_edges.rs, Cargo.toml.

Revision, read-back and maintenance

The relation and token-set paths support MEM-RTE-correction-update and EPI-RTE-relation-withdrawal. They suppress targets without proving either their falsity or replacement truth. Old records remain accessible on alternate paths. The summary selector and its core/service callers support MEM-OBJ-summary-candidate, EPI-OBJ-summary-candidate, MEM-RTE-summary-selection and EPI-RTE-summary-selection: metadata windows, not generated prose. Coverage scans stored local Summarizes edges without requiring a valid summarizer or testing semantic adequacy. SRC-1, src/lib.rs, src/server.rs, src/search/summary_index.rs.

Core lexical seeding, optional graph expansion, bundle accumulation and service pagination support MEM-RTE-bundle-read and EPI-RTE-bundle-selection. The accepted ranked request shape does not turn this into full vector/RRF bundle selection. Path membership is not proposition warrant. The grouped-route descriptions do not claim complete cross-chain expansion correctness. SRC-1, src/lib.rs, src/server.rs, src/search/bundle.rs.

I traced Markdown rendering and import, dashboard agent copy and core deregistration. MEM-OBJ-memory-markdown, MEM-RTE-markdown-exchange, EPI-RTE-markdown-import, MEM-RTE-agent-copy and MEM-RTE-chain-withdrawal retain material exceptions: empty-query export uses the unfiltered renderer; import/copy creates fresh records without original graph exclusions; sequential writes/removals do not guarantee whole-batch rollback or secure erasure. These positive counterpaths justify rejecting universal stale-memory suppression and permanent-history claims. SRC-1, src/lib.rs, src/server.rs, src/dashboard.rs.

RT-OBJ-skill-registry, MEM-OBJ-skill-body, EPI-OBJ-skill-document, RT-RTE-skill-admission and EPI-RTE-skill-withdrawal correctly separate body, version machinery, signature enforcement and lifecycle policy. Upload stores full/delta versions; read reconstructs and exports them. Service signatures are conditional on active keys; dashboard upload reaches the registry directly. Revoked content remains readable with status separate from authored warnings, and subsequent upload does not clear Revoked. These findings do not license safety, quality selection or automatic host rollback. SRC-1, src/skills.rs, src/server.rs, src/dashboard.rs.

RT-RTE-state-maintenance and RT-RTE-release-update retain their administrative scope. Service merge remaps attribution and copies selected fields, losing graph metadata; success removes the source, while append failure can leave a target prefix. Restore checks the archive before staged extraction and fallible target copies; CLI guards do not establish universal process exclusion. Ordinary update prompts interactively, whereas force-update bypasses that confirmation. Cargo installation is not evaluation-driven self-improvement. SRC-1, src/server.rs, src/lib.rs, src/backup.rs, src/cli/mod.rs, src/bin/mentisdb.rs.

MEM-RTE-webhook-delivery is supported as filtered current-event push with bounded queue, concurrency, timeout and retries. The queue can drop events; the external receiver's retention and behavioral force remain unknown. It is not established historical read-back. SRC-1, src/webhooks.rs, src/lib.rs.

Epistemic and workflow dispositions

I checked every epistemic ledger function and its status against the corresponding route, keeping check, admission, retention and consumption separate. MEM-RTE-retrieval-evaluation and EPI-RTE-recall-evaluation are supported by independent reads of all three matchers and aggregation consumers, plus the inspected ingest paths. Bidirectional substring matching licenses the stated metric deductions, including partial/empty-string countercases; ConvoMem abstention is negated retrieval matching; LoCoMo broadens evidence with neighbors. No answer generator or automatic diagnostic-to-successor consumer is established by these routes. SRC-1, lme-benches/longmemeval_bench.py, lme-benches/convomem_bench.py, locomo-benches/locomo_bench.py.

The source wording of EPI-CLM-context-continuity and EPI-CLM-fleet-improvement is faithfully attributed. Selective reload and unchecked summaries do not establish zero loss; instruction versioning and conditional signatures do not establish better skills. The bounded conclusion does not turn missing execution into absent architectural capacity, or conflate learning, reflection, autonomy and improvement. Its negative about evidence-consuming acceptance is explicitly limited to inspected routes, not the whole repository. SRC-1, README.md, src/server.rs, src/skills.rs.

WF-RTE-session-continuation traces one contiguous shipped example. Independent reading confirms that it retains the original handoff index, uses one open chain and checks hit count/backlink rather than executing the proposed auth-test rerun. Core ref validation supports pre-admission rejection; downstream append errors have a different effect. The report appropriately withholds real restart, autonomous delegation, diagnosis acceptance and delivered application repair. The cookbook Handoff-role convention and skill Summary/Checkpoint convention are not claimed interchangeable. SRC-1, docs/cookbook/cookbook-2-2-cross-session-continuity.html, docs/cookbook/cookbook-1-0-operator-playbook.html, MENTISDB_SKILL.md, src/lib.rs.

I checked every Part of relation: content to thought envelope, skill body to registry, correction/withdrawal and workflow handoff to roundtrip, extraction schema to extraction, and lifecycle withdrawal to skill admission. They identify material contained parts rather than replacement candidates. I checked all seven supersessions in Reconciliation: thought content, skill body, summary candidate, relation withdrawal, summary selection, bundles and retrieval evaluation. The paired declarations have the same referents; their different fields and ledger assessments remain attributable to their authors. The broader optional checkpoint progression does not invalidate shared selector identity. Markdown import remains a partial overlap with exchange, not a duplicate complete route. No supplied status or value is rewritten by the amendments. All seven memory integration issues have supported dispositions. No unresolved inter-report conflict or missing required part was identified; source prescription discrepancies and different inspection depths are not fabricated conflicts.

Omission questions came from the frozen top-level tree and shipped producer/consumer dependencies, not just quoted anchors. I asked whether another shipped caller could change the claimed admission envelope, whether compression or dynamic skills close an automatic semantic-update/acceptance loop, and whether evaluators consume diagnostics as production updates. I inspected dashboard copy/upload, service merge, direct versus service renderers, local stdio dispatch, integration-writer searches, evaluator consumers and the additional compression/dynamic-skill cookbook chapters. The alternate admission and loss-of-lineage cases are already preserved. Integration writers configure MCP connections, not evidenced host compliance. The dynamic-skill example authors two versions without an intervening quality evaluation. The compression example adds a documented variant missing from the partial inventory, retained below. SRC-1, src/dashboard.rs, src/server.rs, src/bin/mentisdb.rs, src/integrations/targets/codex.rs, src/integrations/targets/claude_code.rs, src/integrations/targets/opencode.rs, docs/cookbook/cookbook-2-3-dynamic-skills.html, docs/cookbook/cookbook-2-1-semantic-compression.html.

Further optional omission expansion stopped because these selected questions have supported dispositions and no identified material challenge remains unrouted. The reports do not emit exhaustive correctness, security, evaluator or system-wide acceptance-absence findings that would require auditing every available helper, UI default, maintenance instruction or dependency. Their named uninspected alternatives still prevent those comparisons. The existing source/operation uncertainty is faithful scope, not a new defect or evidence of absence.

Blockers

none

Limits

  • The summary/checkpoint inventory is incomplete as a catalogue of shipped compression conventions. Alongside MEM-RTE-summary-selection, MEM-OBJ-thought-content and RT-BAP-core-skill, the cookbook supplies a caller-authored Summary with Compression role and positional refs, followed by role-filtered retrieval. Those refs become References edges through ordinary append, not the Summarizes edges used by candidate-coverage suppression. Withhold complete comparisons of compression forms and their coverage bookkeeping; the supported distinction between read-only candidate selection, separate caller authorship and later retrieval remains valid. This is doctrine/example evidence, not observed automatic summarization or host eviction. SRC-1, docs/cookbook/cookbook-2-1-semantic-compression.html, src/lib.rs. The example's material construction is:

    .with_role(ThoughtRole::Compression) .with_refs(source_indices) --- docs/cookbook/cookbook-2-1-semantic-compression.html