ByteRover Delivers 92.2% Agent Memory Accuracy
ByteRover uses curated knowledge trees and tiered retrieval to achieve 92.2% accuracy on LoCoMo benchmark, outperforming vector stores for portable, local-first AI agent memory.
Curated Tree Structure Replaces Vector Stores
ByteRover builds stateful memory as a hierarchical knowledge tree in natural language format, enabling agents and humans to reason over it like a system. It applies curation to organize content, replacing vector-based retrieval with a tiered pipeline: fuzzy text search first, then LLM-driven deep search for precision. This yields 92.2% retrieval accuracy on the LoCoMo long-context memory benchmark, topping the leaderboard and beating major systems. Import existing markdown or text files (e.g., MEMORY.md) via brv curate -f ~/notes/MEMORY.md or folders, keeping your setup alongside.
Local-First Portability Across Tools
Runs entirely locally by default—no account, cloud, or telemetry required. Push to ByteRover Cloud for version control, editing, sharing with teammates, or use across machines/agents. Memory persists and shares across OpenClaw agents, works with any LLM/provider via API keys, and ports between tools like OpenClaw, Claude Code, Cursor. Enterprise Cloud adds SOC 2 Type II, AES-256 encryption, TLS 1.2+, RBAC, and data residency.
4-Step Setup for Immediate Use
Install via curl -fsSL https://byterover.dev/install.sh | sh (or npm/Windows). Then: 1) Configure LLM/providers; 2) Connect agent connectors; 3) Curate memory (/curate); 4) Query it back (/query); optionally push to cloud. Integrates with OpenClaw for persistent shared memory; full migration guide for existing systems.