COMBINATION MACHINES Towards energy-efficient general intelligence
New Category of AI Models

AI that scales on storage, not compute

PAM Programmable Action Models

PAM is a CPU-based, compute-efficient AI model that learns in real time without requiring model training. It brings software-like capabilities directly into the model: deterministic workflows, native memory, and data management. This can help enterprises build workflows, analytics, and systems without month-long software and AI development cycles, token-based billing, or associated data risks.

4.5 years of research and development completed
PAM / STRUCTURED MODEL NETWORK
The cost structure

Today’s stack separates intelligence from the work it needs to do.

Software, databases, agents and LLM calls are assembled as separate components. PAM is designed to bring those capabilities together in one model.

Typical LLM + agent stack
Pay to build it. Pay to run it.
  • Complex agent setup and software integration
  • Ongoing LLM token costs
  • Separate software, data and agent components
PAM architecture
The model does more of the work.
  • Workflows and data management in the model
  • Runs on standard CPU servers
  • No token-based costs
How it works

Software-like capability, built into the model.

PAM is not LLM- or neural-network-based. It learns incrementally, with deterministic behavior and native memory—without expensive model training.

01 / LEARN

Incremental learning

New knowledge can be added a concept at a time, from simple to complex, rather than relying on expensive model training.

02 / REMEMBER

Native memory

Memory and data management are part of the architecture, not a separate system the model must call into.

03 / ACT

Deterministic behavior

Workflows run as model capabilities. Updates can be expressed as simple instructions instead of month-long software projects.

A different software model

Millions of trade records. The application runs in PAM.

We built an intelligent trade-management system with millions of records in PAM itself. PAM runs the application rather than relying on software plus a separate database and agents.

When workflows change, updates can be given as instructions—not scoped into month-long software projects.

Where we are
4.5 years Research and development completed.
Now Initiating design partnerships while building the enterprise platform.
In development Research & Engineering whitepaper coming soon.