QUEST: A Fully Open Recipe for Training Deep Research Agents from Scratch

Ask the agent
QUEST: What I can research for you?
Try examples
QUEST can handle multiple types of queries as shown below.
Output
Settings
Memory Strategy
Condenser (default) — when context grows large, a State Summarizer LLM compresses earlier turns into a structured JSON of trusted/untrusted/uncertain claims, visited sources, and prior search queries; the agent continues with that compact state.
Vanilla — memory management disabled; the full conversation history is kept.
Discard-all — when context grows large, the entire message history is reset, restarting the agent from the original question with no accumulated context.
Hide-tool-result — when context grows large, older tool responses are pruned; only the most recent tool result is kept.
2 50
0 1.5

QUEST is a fully open recipe for training deep research agents from scratch — covering data synthesis, memory management, infrastructure, and long-horizon training.