Memory Architecture Quality Standard for LLM Assistants

1.1 (2026-08-21) version of the Memory Architecture Quality Standard applies to any long-term memory-enabled LLM assistants and agents. The audit checklist is divided into sections A-K, and must be completed in order, with each item marked as yes, partial, no or n/a. Evidence must be recorded for all nos and partials.

Section 0 establishes that a complete storage map must be compiled before proceeding with other sections. Critical items include identifying all modules that write to or read from each store, confirming write key matches read key, identifying channels returning model output to memory input, and verifying the analyzed code matches executed code.

Section A focuses on input validation, with critical items requiring trust filters for external sources, exclusion of secrets from memory indexing and vectorization, and flagging service content to exclude from context assembly. Provenance of each record is also required, with untrusted records receiving lowered weight or placed in a separate trust zone.

Section B covers write idempotency and integrity, with critical requirements including idempotent writing, correct deduplication, consistent writing to multiple stores, and non-silent write failures. Compressed representation and original data must be written consistently, with write failures logged and retried.

Section C deals with growth management, including defining limits and retention policies for each store, ensuring idempotent re-indexing, defining quotas for single record size and total volume, and monitoring volume dynamics for anomalies. Multimodal input deduplication is also addressed.

Section D concerns retrieval ranking and gating, with critical items requiring context-adaptive scoring, soft relevance thresholds, and preventing unbounded growth of multimodal input hashes. The audit must be completed before issuing the Audit Result Protocol.