MAQS — open audit checklist for verbatim + hybrid memory systems

[alexk202]

Hi Milla and Ben,

MemPalace's verbatim storage model — keeping memories exactly as stated rather than extracting or compressing — is architecturally distinct from everything else in the ecosystem. It's also the model most vulnerable to a specific failure class: storage that scales but retrieval that doesn't.

We've published MAQS, an 82-item audit checklist for LLM systems with persistent memory, built from real incidents across 12 deployed assistants. For verbatim + hybrid systems like MemPalace, three items are especially relevant:

• B1 — concept ID must be deterministic: the same content re-indexed must produce the same ID. Non-deterministic IDs silently multiply entries — in verbatim storage this compounds fast. We saw a creative archive grow from 3,119 to 24,336 entries purely from non-idempotent indexing.
• D6 — hybrid retrieval weight balance: when switching between raw and hybrid modes (as MemPalace does between 96.6% raw and 100% hybrid), retrieval weights must be explicit and configurable — not hardcoded. Silent weight shifts change recall behavior without warning.
• E10 — anchor memories: in a verbatim system, some entries must survive pruning and decay unconditionally. Without protected anchors, long-term identity degrades silently under volume pressure.

The full standard, a completed real-world audit report (15 critical failures to full conformance in 8 waves), and a Russian translation are on GitHub:
https://github.com/alexk202/memory-architecture-standard

CC BY-SA 4.0. If MemPalace runs the checklist — we'd be very interested in what applies and what's missing for a verbatim architecture.

Best regards,
Aleksandr Kossarev
Jõgeva, Estonia
GitHub: https://github.com/alexk202
Dev.to: https://dev.to/aleksandr_kossarev_e23623