arXiv: ByteRover — Agent-Native Hierarchical Memory Architecture Hits 92.8% on LongMemEval With 98.7% Knowledge Update Accuracy

arXiv·medium signal

ByteRover (2604.01599) proposes an agent-native memory system where the same LLM that reasons about a task curates knowledge into a hierarchical Context Tree. Instead of treating memory as separate infrastructure, the agent organizes, compresses, and retrieves its own context. Results: 92.8% overall on LongMemEval, 98.7% on knowledge update, 91.7% on temporal reasoning, and 96.7% on single-session preference. Directly relevant to builders designing persistent agent memory systems.