KVMs - KV Cache Aware Models

Training language models that share what they know, not just what they say.

Overview

AI systems are increasingly built from many models working together: planners handing off to executors, specialists consulting one another, and agents that review, critique, and build on each other's work. Today, almost all of that collaboration happens through plain text. One model writes out its thoughts, and the next model reads them from scratch.

That works, but it is wasteful. Every time a model reads a message, it rebuilds an internal understanding of that context that another model has already built. KVMs explore a different approach: models trained from the start to pass along that internal understanding directly, so agents can pick up where another left off instead of starting over.

The idea

As a language model reads, it builds up a working memory of everything it has seen so far, often called the KV cache. Normally this memory is private to a single model and thrown away when a conversation ends. It was never designed to be shared.

KV cache aware models are trained with sharing in mind. Rather than treating that memory as a temporary byproduct, a KVM learns to produce memory that other agents can understand and use, and to make good use of memory handed to it by others. The goal is for collaboration between models to feel less like passing notes and more like a shared train of thought.

Why it matters

• Speed. Agents can skip re-reading long contexts that another agent has already processed.
• Cost. Less repeated work means less compute spent per task, especially in large multi-agent pipelines.
• Richer handoffs. Internal memory can carry nuance that is easily lost when everything has to be squeezed into text.
• Scale. As systems grow to dozens or hundreds of cooperating agents, efficient communication becomes a bottleneck worth solving.

Built for multi-agent setups

Most models are trained to be excellent on their own, then wired together after the fact. KVMs flip that around. Training focuses on the situations where models actually collaborate: handing off a partially completed task, splitting a large problem among specialists, or bringing a new agent up to speed on a long-running conversation.

The result is a family of models that treat teamwork as a first-class skill. They are meant to be good teammates, not just good individuals, and to stay reliable when the context they receive comes from another model rather than from a person.

Open questions

This is an active area of exploration, and there is a lot still to figure out:

• How much can be shared between models of different sizes or architectures?
• What is the right balance between sharing internal memory and communicating in plain language?
• How do we keep shared memory interpretable and trustworthy?
• What new kinds of multi-agent systems become practical once handoffs are cheap?

Status

KVMs are a work in progress. More details, results, and write-ups will be shared here as the project develops.