Memori reconciliation

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

Reconciliation

Boundary and connection basis

The reports share the revision and partial-loop scope in boundary. The shipped clients, native engine, integrations and evaluators are included; hosted extraction, admission and compaction internals remain unavailable. The Source register supplies implementation, doctrine/design and reported-operation evidence, not an observed run or causal experiment. This reconciliation uses the supplied members only; no source rereads were needed. No analyst correction answers or refusal were supplied for this attempt.

The amendments below identify duplicate referents, not replacement findings. Both declarations and their analyst-owned fields remain available. A memory declaration's more detailed storage account does not overwrite an epistemic declaration's warrant assessment, and the epistemic ledger's implemented classification does not supply observed operation to another member.

Duplicate supersessions

Amendment: EPI-OBJ-entity-facts is superseded by MEM-OBJ-entity-fact. Both identify memori_entity_fact.content as retained natural-language entity assertions, distinct from the transient recall envelope. Identity evidence is the epistemic declaration's Memories.entity.facts mapping and the memory declaration's Python response-to-fact-row lineage, supported by their parser and SQLite passages. This connects the epistemic inventory and service-import, scheduling, retention and recall findings to the memory account's same fact content. It preserves the epistemic source-warrant unknown and the memory account's native mock/fallback qualifications; neither becomes a truth certificate or proof of adapter persistence.

Amendment: EPI-OBJ-semantic-triples is superseded by MEM-OBJ-knowledge-graph. Both name semantic triples admitted to memori_knowledge_graph, with subject/predicate/object and type fields. Identity evidence is the epistemic triple parser account and the memory declaration's quoted knowledge_graph.create mapping. The affected epistemic shape-check, rendering, retention and withdrawal findings concern these same structured assertions. This does not merge graph rows with their separately rendered fact text or establish a later graph-query-to-model consumer.

Amendment: EPI-OBJ-conversation-summary is superseded by MEM-OBJ-conversation-summary. Both identify Memories.conversation.summary mapped to memori_conversation.summary, overwritten by update and reused in augmentation and recall. Identity evidence includes the epistemic scheduling passage for conversation.update and the memory declaration's SQLite update and latest-message selection passages. Inventory, import, scheduling, retention and recall findings retain their own assessments of this same stored summary. The memory report's Python-versus-Rust summary-input distinction remains local; no universal recursive-summary lineage follows.

Amendment: EPI-OBJ-process-attributes is superseded by MEM-OBJ-process-attribute. Both identify returned process.attributes admitted to process attribute rows. Identity evidence is the epistemic parser passage extending attributes and the memory passage scheduling process_attribute.create. This connects the epistemic inventory and service-import, scheduling and retention findings to the memory storage account. Both leave later model-facing consumption uninspected; possible policy content does not establish delivered skills or operative improvement.

Amendment: EPI-OBJ-fact-embeddings is superseded by MEM-OBJ-content-embedding. Both identify fact_embeddings / content_embedding as stored dense numerical representations used by fixed ranking rules. Identity evidence is the epistemic parser/search account and the memory declaration's vector attachment, missing-vector and retrieval account. This connects the epistemic inventory and recall-selection inputs to the same retained vectors. It neither equates vectors with producing model weights nor supplies missing immutable model provenance. The lab FAISS index remains a separate artifact with external input provenance.

Amendment: EPI-RTE-recall-selection is superseded by MEM-RTE-fact-selection. Both isolate query/entity-gated Python fact selection and thresholded fact/summary delivery inside the registered invocation. Identity evidence is their common weighted cosine/lexical ranking and injection-threshold account, with summary selection following the selected facts. The memory declaration adds native ranking coverage; the epistemic declaration adds cloud-string scoring and the limit on epistemic endorsement. This connects the ledger's operational selection function to the memory read-back route without promoting either route's operation status. Their shared parent remains RT-RTE-python-invocation, not another duplicate of the selection subroute.

Amendment: EPI-RTE-entity-withdrawal is superseded by MEM-RTE-entity-withdrawal. Both trace the public BYODB deletion call through entity resolution to graph deletion followed by fact deletion. Their identical quoted deletion calls and preserve-conversations findings establish identity. The affected epistemic ledger and bounded conclusion concern the same maintenance operation. Both retain replayable conversations, already delivered context, concurrency and re-derivation limits; neither supports global forgetting or truth-directed rejection.

Amendment: EPI-RTE-feedback-delivery is superseded by MEM-RTE-feedback-submission. Both trace foreground-agent/user feedback through the shipped CLI or host tools to the hosted feedback endpoint, returning success/error without an inspected local adaptation consumer. Identity evidence is the epistemic CLI POST passage and the memory OpenClaw agentFeedback passage; these are entry surfaces of the same submission operation, not separate capture or correction mechanisms. The ledger's transport function and memory integration issue remain bounded to dispatch. Neither declaration establishes feedback-driven memory revision, policy change or learning.

The conversation declarations have overlapping but unequal scope. MEM-OBJ-conversation-message identifies persisted role/content rows and their lossy replay. EPI-OBJ-conversation-messages also covers completed-turn trace transmission, tool IDs and arguments across hosts. A whole-record identity and supersession are ruled out: trace transport is not the same retained row shape, and its private retention remains unresolved. The stored-row object can coexist with that broader capture/history account. Neither record warrants lossless tool-trace restoration. RT-OBJ-recall-context is different from both: it is the transient selected display text supplied to inference, not captured history or a durable fact, summary, vector or access structure.

The epistemic function records EPI-RTE-service-import, EPI-RTE-triple-shape, EPI-RTE-triple-rendering, EPI-RTE-memory-scheduling and EPI-RTE-local-retention are declared parts of RT-RTE-augmentation-admission. They distinguish acquisition, structural checking, rendering, operational disposition and retention within that route. Identity with the whole runtime route is ruled out. The runtime container remains valid; none of these parts replaces all of its dispatch, native admission or recovery findings. In particular, the scheduling function's ledger category does not assert evidence-consuming truth acceptance.

Likewise, EPI-RTE-capture-cleanup is a transformation within RT-RTE-openclaw-hooks, not another declaration of that lifecycle. Its fence-truncation deduction makes the runtime's bounded parsing/cleaning account more specific. It does not contradict a claimed lossless runtime route: the runtime record makes no such guarantee.

The benchmark records form a valid container and function-level account rather than duplicates requiring supersession. MEM-RTE-benchmark-reload spans imported data, index construction/reload, selection, answering, judging and reporting. EPI-RTE-benchmark-answer, EPI-RTE-benchmark-judging, EPI-RTE-benchmark-metrics and EPI-RTE-benchmark-retention distinguish consequential functions inside that sequence. Whole-route identity is ruled out because no one function has the complete reload route's endpoints. The container's findings remain valid when these functions are separated.

MEM-OBJ-benchmark-artifacts inventories index, metadata, answers and judgments with their different consumers. EPI-OBJ-benchmark-answer and EPI-OBJ-benchmark-verdict isolate generated answer and judgment parts; neither supersedes the valid artifact inventory. EPI-CMP-benchmark-embedder and EPI-CMP-benchmark-inference identify lab configurations, not duplicate production components RT-CMP-sentence-transformers-embedder and RT-CMP-provider-model. Shared model families or API use do not establish component identity across these call sites.

The workflow declares no new records. Its annotation develops RT-RTE-claude-skill, rather than requiring a supersession. The ambient-conduct claim RT-CLM-ambient-lifecycle is distinct from the SDK/service operation claim EPI-CLM-structured-memory, the benchmark warrant claim EPI-CLM-benchmark-warrant and the improvement claim EPI-CLM-agent-learning. They concern different claimed operations and evidence requirements.

Integration-issue dispositions and independent convergence

Native alternatives qualify service dependency, not the frozen boundary. The memory report's integration issues identify mock response and empty-operation/raw-content fallback emission in MEM-OBJ-entity-fact. RT-RTE-augmentation-admission describes the normal service-derived path; its runtime account explicitly bounds native service dependence to non-mock augmentation. The epistemic service-import record is Python-scoped. Thus these records do not disagree about every possible native operation: fallback batch emission is not established adapter persistence, and normal local storage remains compatible with remote derivation. No service-approved status is transferred to mock or fallback content.

Graph and process consumers remain a shared limit. The graph and process-attribute declarations above establish storage, while the memory integration issues leave independent later model-facing readers uninspected. The epistemic inventory and retention account retain the same limitation. This prevents extending fact-context delivery to all stored outputs; it does not negate the separately wired fact/summary route.

Hosted memory and priority remain unresolved. MEM-OBJ-agent-memory identifies opaque retained/output representations, not a duplicate of RT-BAP-host-memory-guidance. The latter identifies the consumer, channel and requested force. Both distinguish advertised compaction fields from inspected actual payloads and requested user priority from enforcement. RT-RTE-hermes-provider, RT-RTE-claude-skill, RT-RTE-openclaw-hooks and RT-RTE-mcp-guidance retain different selector, capture and admission policies. Claude's ambient capture and MCP's durable-only guidance are different surfaces, not conflicting descriptions of one unconditional policy.

Completion and delivery limits converge across lenses. The memory augmentation annotation and EPI-RTE-local-retention connect internal SQLite commits and outer rollback limits to the runtime's best-effort completion account. The workflow account independently traces conversation-turn acceptance followed by non-fatal collector failure in RT-RTE-claude-skill. Success JSON, queue drain, provider return, user-task delivery and durable derived memory remain different boundaries. None of these reports claims observed failure or verified end-to-end success. RT-RTE-stream-exhaustion remains a separate persistence route, not evidence that every completion dispatches augmentation.

Benchmark checking does not contradict the open-traffic oracle limit. The runtime route records describe no answer oracle for normal capture/admission. The memory benchmark route and epistemic judging/metrics records establish a separate bounded gold-answer evaluator. They agree that reports are not installed as production memory updates. The epistemic parser countercase, denominator restriction and public token-reduction inconsistency qualify benchmark warrant; the memory report adopts no numerical result that depends on denying those limits. The numerical inconsistency is internal to the public claims recorded in EPI-CLM-benchmark-warrant, not a disagreement between analysts.

Retained influence is not demonstrated improvement. Runtime context delivery, memory lineage/read-back and EPI-CLM-agent-learning converge on a possible later behavioral channel while keeping exercise, benefit and attribution unestablished. No conclusion about model-weight change, reflection, theory criticism, autonomy or self-improvement is borrowed from another property. Their uncertainties remain separate rather than becoming a single negative verdict.

No unresolved cross-report disagreement was identified in these supplied versions. The amendments connect duplicate identities; containment and scope distinctions rule out the other apparent identities. The stated service, consumer, execution, adapter and causal gaps remain limitations, not settled conflicts or corrections to analyst findings.