Reverie is built by Mudit Sarda.
By day, Mudit is a software engineer at Dhiway in Bangalore, building trust infrastructure, which mostly means teaching software to say “prove it”. An engineer by training, Mudit has coded with AI for a year and a half: chat windows, then IDE agents, then agents in the terminal.
Why we built it
One evening, halfway through a session, a permissions error broke the terminal. The way out was a fresh terminal and copy-pasting every prompt and answer from the old one. It worked. But what about next time, and the time after that? A developer gets better with every session; the agent starts each one from zero, however many times you’ve explained things. We built Reverie so what you work out with an agent stays with you, in the next session and with the next agent.
How we work
Understanding over information. Reverie keeps the conclusions an agent reaches, not a record of everything it did, because understanding is what carries over from one session to the next.
Measure, then believe. We benchmark our own ideas, and the first one failed: our first version loaded memory before the agent began work, and on our benchmark the agent did worse. Negative results matter, so the runs are on the EC-Bench page .
Memory you can inspect. Memory is stored locally, you review what is kept, and conclusions are grounded in your repository. Superseded documents stay available, labelled with what replaced them.
Names
Reverie is the product. Engineering Cognition is the research program. The code uses the prefix ec. EMS was the first version.