MerchantBench memory profile

Type: agentic-system-analyses/types/agentic-system-memory-profile.md

Comparison rationale

The local memory mechanism is optional full-document authoring and explicit recall, alongside process-local historical context and a separately persisted inspection trace. The document is not automatically reintroduced after eviction. A requested return is pull; automatic reuse of the surviving historical suffix is a separate coarse push operation. The archive's only established later consumer counts version markers. That computation establishes no routing authority: it constructs output numbering without selecting an operation, destination, scope or input window. The profile does not invent a model restore route (MEM-RTE-document-read, MEM-OBJ-history, MEM-OBJ-memory-history, RT-RTE-context-trim).

The optional experience-to-note path supports afforded trace learning, not a wired summarizer, mandatory successful save or observed improvement. It does not establish consolidation: the reminder requests preservation, while the write handler accepts arbitrary full bodies. The semantic transformation remains undetermined without content, as retained in EPI-RTE-memory-document. The implemented replacement and eviction findings remain separate from this unresolved step. A model choosing its own content followed by software admission is automatic agency; the human editor's explicit save is a separate manual route. The update consumer's learning force is distinct from the resulting note's knowledge force at later model calls. Plans in a requested tool result do not thereby become standing instructions. No semantic admission criterion follows from the byte limit (MEM-RTE-document-write, MEM-BAP-document-guidance, RT-RTE-context-trim).

The trace-source findings describe original conversational and tool content available to that optional write. No independent event-stream origin is inferred from simulation timestamps or JSON envelopes. No trajectory-training/update route is established merely by sequential logs. Raw recording and suffix deletion do not on their own produce a durable derived continuation summary (MEM-OBJ-history, MEM-OBJ-step-traces, MEM-CLM-telemetry-limits).

Reconciliation supersedes the epistemic scratchpad object with MEM-OBJ-memory-doc. All document classifications use that canonical object. The combined epistemic admission/retention/recall container remains valid; the more specific memory routes provide the local warrants here. No supersession changes a controlled value or upgrades its basis.

Every axis remains partial because actual external Hermes memory transformations and consumers are uninspected. Visible filesystem setup, SQLite inspection and readable session display do not classify opaque consumed payloads, establish skill learning or exclude alternate native memory. Known local findings are not weakened or promoted by that uncertainty (RT-RTE-hermes-launch, MEM-RTE-trace-inspection, MEM-CLM-telemetry-limits).

Source-read log: none. This profile classifies supplied accepted records and annotations after reconciliation; it adds no source evidence or record declarations. All positive findings are implementation wiring or explicitly marked affordances, not observed operation or causal benefit.