MemoryBank: Hierarchical Memory for Scalable Long-Horizon Agentic Workflows
New memory architecture enables agents to maintain coherence over 100+ step workflows by compressing episodic traces into queryable knowledge graphs.
Berkeley’s MemoryBank solves the agent amnesia crisis. Agents forget critical context mid-task because KV caches bloat to gigabytes. This paper’s hierarchical approach distills 100-step traces into 10x smaller graph summaries, letting agents ‘remember’ why they chose Tool A over B three hours ago.
What changed. From dumb storage to smart recall: agents now query causal histories (“Why did pricing fail last Q3?”) with 92% accuracy vs 47% RAG.
Why it matters. Production agents processing 1M+ journal entries (shoutout EY’s recent rollout) need this to avoid catastrophic drift.
Builder takeaway. The GitHub repo integrates with LangGraph in 47 lines—test on your longest workflow today. Read the paper.