Y Combinator-backed startup Theta Software, which builds self-learning and real-time adaptation for AI agents, has recently launched. The startup claims it is starting with an intelligent memory layer so that agents can remember and learn from previous interactions. This memory layer uses real-time reinforcement learning (RL) to analyze every run for mistakes and optimizations with a simple four-line addition to existing code.
According to Theta, agents struggle to adapt to complex, real-world workflows. Workflows are dynamic, but agents remain static. They get stuck in loops and require constant human guidance. Even when an agent manages to complete a task, this long process starts all over again for any future run. Learning is fundamentally iterative, but agents can’t learn because they have no memory across runs.
READ: OpenAI unveils Sora, an AI tool that transforms text Into video (December 11, 2024)
Part of Spring 2025 cohort of Y Combinator, Theta claims it has been able to improve the accuracy of OpenAI Operator by 43% with the memory layer. With optimized trajectories, Operator also took seven times fewer steps, resulting in better speed and cost.
Theta has been founded by Rayan Garg, Tanmay Sharma, and Gurvir Singh. While Garg has previously done machine learning (ML) research as head of product at DeepSilicon, Sharma has developed browser agents at MultiOn. Singh’s previous experiences include building distributed post-training systems at Cornell.
Theta attempts to resolve this problem by building the infrastructure for agents to self-learn and adapt in real-time. By adding four lines of extra code, Theta creates an intelligent memory layer that learns from the agent’s previous runs. At the end of a run, Theta generates and embeds an analysis of your agent’s trajectory, identifying critical steps and assessing overall performance. Before the next run, it generates a detailed plan based on the specified task. It gathers relevant insights and optimizations based on previous runs.
With real-time RL, the memory layer continues to improve with more runs, meaning the agent gets better and better without any human intervention.