HMS: A Memory Protocol Built for AI

新智元 · wechat · 2026-07-15

This in-depth article introduces a memory system named HMS (Holographic Memory System). Its core goal is to decouple "memory" from context windows and RAG, transforming it into a cross-session, inheritable, and verifiable long-term memory layer. Drawing from cognitive science, the article distinguishes between retention and recall: retention handles compression and consolidation, while recall reconstructs the chain of evidence based on cues. HMS further breaks down recall into three steps: query parsing, on-demand retrieval, and result self-checking. It provides six types of retrieval "keys" to handle time, people, context, latest values, contradictions, and relationship jumps. The author claims this mechanism achieves SOTA on both the LongMemEval and LoCoMo long-term memory benchmarks, leading in sub-tasks like temporal reasoning, multi-session aggregation, and multi-hop reasoning. The latter half emphasizes that this is a "memory protocol" rather than a single product. Developers can integrate via an SDK, while everyday users can manage their personal memory accounts using MemoryBank, supporting local mode, authorization, auditing, hiding, and deletion. The author compares it to foundational infrastructure like an OS, TCP/IP, or CUDA in the AI era, predicting that the memory layer will become a shared common foundation across different models, Agents, and robots.

Related event: AI Memory is Evolving into an Independent Infrastructure Layer(2 posts)→

Original post →

More from Research

Research channel →