AI agents are generating detailed records of the decisions they make, the tools they use and the outcomes they produce as they take on multi-step tasks. Researchers at Google and Microsoft are exploring ways to turn those records into reusable knowledge, procedures or testing environments. Google’s Dream-RSI research converts the history of earlier discovery processes into a replayable simulator. New exploration strategies can then be tested against past results without the agent repeating the original work. In experiments the approach reduced discovery-agent calls by up to 162 times while maintaining or improving quality in several settings. Microsoft’s PlugMem research extracts facts and reusable skills from interaction histories and organises them into structured memory. Its Foundry Agent Service is adding procedural memory so agents retain successful execution patterns rather than only facts. Early evaluations showed about a 5 percent improvement on selected benchmarks.
AI agents are generating detailed records of the decisions they make, the tools they use and the outcomes they produce as they take on multi-step tasks. Researchers at Google and Microsoft are exploring ways to turn those records into reusable knowledge, procedures or testing environments. Google’s Dream-RSI research converts the history of earlier discovery processes into a replayable simulator. New exploration strategies can then be tested against past results without the agent repeating the original work. In experiments the approach reduced discovery-agent calls by up to 162 times while maintaining or improving quality in several settings. Microsoft’s PlugMem research extracts facts and reusable skills from interaction histories and organises them into structured memory. Its Foundry Agent Service is adding procedural memory so agents retain successful execution patterns rather than only facts. Early evaluations showed about a 5 percent improvement on selected benchmarks.