Mem0 profile verification

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

Verification

I checked all ten axes in the memory profile against the handed runtime, memory, epistemic and workflow findings, resolving record identity through the supplied reconciliation. This included positive values, their evidence bases, local assessments, natural units and unresolved inventory coverage. The payload supersession resolves EPI-OBJ-data to MEM-OBJ-memory-data; the profile preserves the distinct history, keyword, vector and route referents. The previous verification and the author's correction answer were assessed as arguments, not accepted judgments.

The prior consolidation objection is resolved. The local Python extraction/procedural unit now cites only RT-RTE-extract-add and MEM-OBJ-procedural-memory for wired consolidation. The separate hosted checkpoint-derivation unit cites MEM-RTE-coding-flush, emits no operation and is not-determinable. Its note and the rationale distinguish implemented selection/redaction/submission from the unavailable hosted curation transformation. Neither the remaining curation findings nor the trace-learning finding depends on assigning that hosted extractor a wired consolidation operation: the latter classifies a trace-fed write/later-consumer route, not the remote transformation's semantic operation.

Storage and representational form preserve operative-part distinctions. Vector text/access/embedding/entity parts, SQLite messages/history/evidence, persisted configuration and session-local seen IDs have recorded substrates. Service-object is an afforded interface boundary, not a claim about remote physical storage. Text, symbolic fields and dense representations remain separate; hosted display text does not classify the opaque payload. See MEM-OBJ-memory-vectors, MEM-OBJ-entity-links, MEM-OBJ-evidence-store, RT-OBJ-config-overrides, MEM-RTE-pi-memory and RT-RTE-hosted-client.

Lineage and write agency retain the distinction between derivation, original authorship and admission control. Conversation-derived facts/summaries, imported caller/file content and computed access representations have positive witnesses. Unknown initial content and encoder provenance are not resolved by those witnesses. Software/model extraction, capture and bookkeeping establish automatic admission; generic direct mutation, import and configuration initiation do not establish a per-content human decision. See RT-RTE-extract-add, RT-RTE-direct-write, EPI-RTE-cli-import, RT-RTE-runtime-config and MEM-RTE-hermes-memory.

Behavioral authority is supported at named consumers and effects: recalled facts supply knowledge; installed extraction guidance supplies instruction; retained experience/context influences later fact-producing writes; stored representations influence ranking; scopes/checkpoint state route selection and submission; retained hashes and seen IDs supply eligibility criteria and enforcing exclusions. These do not establish truth acceptance, universal suppression, model compliance or improved capacity. See RT-BAP-answer-context, RT-BAP-extraction-context, RT-OBJ-custom-instructions, EPI-RTE-extraction-checks, EPI-RTE-memory-selection and MEM-RTE-coding-recall.

The remaining curation findings distinguish bounded duplicate suppression, ID-preserving evolution, withdrawal with surviving history and decay through tail eviction/default expiry visibility. Their notes preserve direct-ID/extraction counterpaths and non-atomic persistence. The cessation inference supports claimed synthesis only; it does not establish observed novel production or universal semantic fidelity. See EPI-RTE-extraction-checks, MEM-RTE-typescript-memory, RT-RTE-direct-write, MEM-OBJ-history, MEM-OBJ-messages and EPI-CLM-evidence-bound.

Read-back direction separates requested backend retrieval from automatic host-context supply. Read-back signal classifies the actual push selectors: scope/count/budget rules supply coarse selection, accumulated seen-ID matching selects unseen delivered facts, and parent-session identity selects stored recall for sidekick delivery. Backend relevance algorithms remain unresolved rather than inferred from query strings. The incompatible proxy is not counted as successful recall delivery. See RT-RTE-openclaw-hooks, MEM-RTE-coding-recall, MEM-RTE-pi-memory, MEM-RTE-hermes-memory and RT-RTE-proxy-chat.

Trace learning requires automatic trace-fed retained results and later consumers, not merely raw-log storage. The local facts/continuation summaries and shipped capture-to-recall chains support the positive route findings without implying trained weights, exercised improvement or hosted compliance. Trace source follows original content: conversational turns support session-logs and selected execution evidence supports tool-traces. Static repository material, event wrapping and multi-turn serialization do not add event-streams or trajectories. Procedural original-input provenance remains explicitly unresolved. See RT-RTE-chat-memories, MEM-OBJ-procedural-memory, MEM-RTE-coding-flush and MEM-RTE-coding-recall.

The profile groups some local parts/routes with shared values and coverage rather than providing a fine-grained census. These groupings retain the consequential distinctions above and do not upgrade opaque alternatives. Every axis remains partial; hosted internals, alternative stores, other adapters and unread consumers remain included and unresolved. No emitted axis asserts absence, inapplicability or complete system coverage. The remaining uncertainty is faithfully scoped, not an additional publication defect. No observed or causal comparison basis is emitted.

Source-read log: none. This verification uses only the handed records and their evidence; it adds no replacement profile, records, quotations or execution evidence.

Blockers

none

Limits

none