XNews

Agent memory & retrieval systems (MemOS, RAG, DeepSeek research)

Agent memory & retrieval systems (MemOS, RAG, DeepSeek research)

MemOS 2.0 (Stardust) presents a memory-first OS for agents; DeepSeek papers propose engram–MOE balance for memory and reasoning.

Key insights

MemOS 2.0 introduces OS-like memory management for agents

MemOS 2.0 (Stardust) treats memory as a system resource with MemCube units combining data and metadata (source, version, permissions, forgetting policy). It implements a three-tier storage: working memory mapped to KV cache, long-term vector/graph stores, and parameter memory via LoRA-style interfaces. A scheduler decides which memories to activate per interaction to reduce latency and token costs and to improve coherence across multi-session agents.

DeepSeek research proposes optimal memory–reasoning ratio

DeepSeek papers (authored with Liang Wenfeng) presented 'Engram' concepts and an empirical finding that a 25% memory / 75% reasoning allocation yielded best performance on test tasks; adding a memory dictionary freed compute for deeper reasoning, equating to ~7 extra network layers of effective reasoning depth in experiments.

RAG-from-scratch and MemOS alternatives show varied approaches to private knowledge

Open-source rag-from-scratch tutorials cover indexing, retrieval, routing and adaptive RAG techniques; MemOS proposes a deeper OS-level scheduling approach that aims for millisecond recall and reduced token usage compared with typical RAG implementations.

Claude Cowork knowledge-base upgrades mirror MemOS aims for persistent memory

Reports indicate Anthropic is adding permanent memory / Knowledge Bases to Claude Cowork; community notes MemOS-style projects already implement persistent multi-modal memory with fast recall and token savings, plus natural-language corrections to memory entries.

Source posts