Memory for
Long-Running AI Agents
KarLex AI is building CIEL, an early language model that helps AI agents remember important details without rereading their entire history.
Building CIEL for Agents with Long Histories
CIEL is a language model for agents that work across long conversations, code, tool output, and documents. Today those agents often reread their whole history or rely on summaries that can drop an important detail.
CIEL gives working memory, recent context, and selected older text different jobs. The current model is an early research build, not a finished product. We are testing questions from real documents and faster generation with teams that have long-running agent workflows.
Alternatives require evidence
We study a common question across AI systems: when can an expensive or conventional computation be replaced by an alternative, and what evidence is required to accept—or reject—that substitution? Our work spans independently checkable search, local learning under constrained compute, and exploratory quantum-classical methods.
Three Parts for Three Memory Jobs
Instead of treating every old token the same way, CIEL gives each part of a long history a clear role.
Working Memory
Efficient recurrent layers process most of the history as a compact running state.
Recent Context
Local attention keeps the latest conversation, code, and tool output easy to access.
Older Details
A separate memory layer retrieves selected older text when the agent needs it.
Does Your Agent Need to Remember Old Details?
We are looking for a team with a long-running coding, research, support, or operations agent to test CIEL on a real workflow.