Voaige

Founded 2021

Execute test-time search and episodic memory to improve language model reasoning

You can use Voaige to run test-time cognition for autonomous software agents without changing model weights. The system acts as an inference layer between your agent and language models, coordinating deliberate search trees and episodic memory. When agents encounter difficult steps, it searches candidate paths, evaluates them against evidence, and stores successful trajectories for future tasks. This approach reduces overall compute expenses across complex domains like customer support, programming, IT operations, and financial research while maintaining consistent reasoning performance across multiple model families.

What Voaige does

Coordinate test-time search across multiple reasoning pathsRetrieve past trajectory steps from episodic memory banksEvaluate candidate model outputs against task evidenceSave validated execution paths back to episodic memoryDirect inference calls between agent harnesses and foundation modelsNarrow exploration spaces during multi-step automated agent runsInference layer inserted between agent harness and model weightsDeliberate search trees for complex multi-step reasoning stepsEpisodic memory bank to recall and reuse past problem-solving routesMulti-model coordination across an agent trajectoryCandidate path evaluation and verification against environmental evidenceInference-time compute management to reduce overall token costs