Mem0 memory profile
Type: agentic-system-analyses/types/agentic-system-memory-profile.md
Comparison rationale
The ten axes remain partial. Inspected local objects and shipped consumers establish positive mechanisms, while service internals and unread included alternatives prevent complete coverage. A wired classification describes source-connected operations, not observed execution, compliance, truthful extraction or improved capacity. Scope follows the accepted inventory rather than only the strongest Python witness.
Reconciliation supersedes EPI-OBJ-data with MEM-OBJ-memory-data. All payload findings use that canonical object; its old/new history copies remain MEM-OBJ-history. The symbolic keyword part remains separately citable as EPI-OBJ-text-lemmatized, without replacing the valid access-payload container. No other supplied supersession changes a value.
Dense vectors are retained parametric representations, not the external encoder's weights. Entity links do not establish graph storage. The service-object affordance preserves the opaque hosted alternative rather than exporting local vector/SQLite architecture to it. See MEM-OBJ-memory-vectors, MEM-OBJ-entity-links and RT-RTE-hosted-client.
Inspected local Python fact extraction and procedural summaries qualify as wired consolidation and automatic trace-fed memory writes when their retained results reach later context. The coding-checkpoint branch separately wires selection, redaction, submission and later recall, but its unavailable hosted extractor does not establish a wired curation transformation. Its curation unit therefore retains no positive finding and a not-determinable assessment; requested extraction instructions do not resolve that operation. See MEM-RTE-coding-flush. This distinction leaves the local consolidation positive intact and does not settle instance-level entailment. The prescribed cessation inference supports only claimed synthesis; an extraction error by itself would not establish synthesis. No parameter-training conclusion follows. See RT-RTE-extract-add, MEM-OBJ-procedural-memory, EPI-CLM-evidence-bound and RT-BAP-extraction-context.
Curation is consumer-specific. Hash deduplication is bounded, expiry does not prevent direct ID/extraction reads, and vector withdrawal leaves history and recent messages. Explicit replacement establishes evolve without universal human control. TypeScript ADD-shaped returns can include failed insertions and do not inherit Python result-accounting guarantees. See EPI-RTE-extraction-checks, EPI-RTE-memory-selection, RT-RTE-direct-write and MEM-RTE-typescript-memory.
Read-back separates a backend pull request from automatic host-context push. Query strings alone do not establish a remote embedding, lexical or model-judgment selector. The OpenClaw and coding/Pi/Hermes backend relevance selectors retain their unresolved algorithms. Seen-ID filtering is an actual identity match affecting delivered facts, not merely an ID attached to a requested object. Sidekick delivery separately selects stored parent recall by tracked identity. See RT-RTE-openclaw-hooks, MEM-RTE-coding-recall, MEM-RTE-pi-memory and MEM-RTE-hermes-memory. The incompatible proxy is not counted as successful recall delivery; see RT-RTE-proxy-chat.
Trace origins follow original content: coding tool evidence remains tool-traces; conversational turns remain session-logs even when captured as host events. Static repository evidence supplies no additional trace source. Opaque provenance, initial-content control, alternative stores, remote transformations and unread integration routes prevent stronger assessments. See MEM-RTE-coding-flush, EPI-RTE-cli-import and RT-RTE-runtime-config.
Source-read log: none. This profile uses only supplied accepted records and reconciliation; it adds no records, source quotations or execution evidence.