HiMeS: A Hippocampus-Inspired Agent Memory System for AI Agents
arXiv 2026
A deep engineering dissection of HiMeS, a hippocampus-inspired agent memory system pairing short-term query rewriting with long-term memory-aware reranking.
- Separates memory by pipeline stage: short-term memory rewrites the query before retrieval; long-term memory re-ranks retrieved chunks after it
- Short-term memory is a trained query rewriter — SFT followed by GRPO with the HSER reward, all training-time only
- Long-term memory never enters the prompt; it acts as an attention-inspired relevance signal deciding which retrieved context survives
- Partitioned long-term memory via Atomic Topic Modeling narrows the search space before semantic similarity is computed
- Includes my own engineering interpretation: a three-layer raw / structured / semantic memory design with multi-signal retrieval
Agent Memory
Memory Systems
HiMeS
Long-Term Memory
Short-Term Memory
RAG
Query Rewriting
GRPO
Semantic Retrieval