Building CIEL

Memory for
long-running
AI agents.

We’re building CIEL to help agents carry important context across conversations, code, and decisions.

CIELEarly research

AI research & developmentExplore the research

The idea behind CIEL

Keep the thread.

Useful context lives across conversations, code, tool output, and documents. CIEL explores how to keep the relevant pieces available as the work moves forward.

It is a research prototype. We’re studying three complementary roles for memory, and evaluating them on long histories.

Continuity, across the work.

Working memory

The decisions and constraints that should stay close to the task.

Recent context

The latest conversation, code, and tool output that inform the next step.

Older details

Relevant information from earlier work, brought back when it is needed.

We’re looking for teams building agents that work over time.

Explore CIEL

Ideas & experiments

Inside our research.

All publications

4 current manuscriptsUnder review · September 2026

Transformer learning

How should a Transformer learn?

What happens when a Transformer learns in separate stages? Two CPU studies examine the speed, memory, and learning quality of local training and predictive coding.

Machine Learning