The last 3 years of AI progress didn't come from the AI models. But from what was moved out of them.
That's the finding of a review by an international team of 24 researchers: every leap in AI systems that actually work came from what engineers moved out of the model. Namely: memory, skills and protocols.
They call it "externalization": relocating the cognitive burden into persistent, inspectable, reusable structures.
In three years, the industry has moved through three AI eras:
1/ 2022-2023: capability lived in the model's weights. Whoever had the best model won.
2/ 2023-2024: capability lived in the context. Whoever wrote the best prompts won.
3/ Since 2024: capability lives in the infrastructure around the model: the harness. Whoever builds the best one wins.
Where is your company on that timeline?
The research says three components must move out of the model. And they translate directly into organizational assets:
1/ Memory: your decisions, results and learnings, persisted where the AI can find them. The paper's elegant point: this turns a recall problem into a recognition problem. Writing did the same for human cognition.
2/ Skills: your ways of working, packaged as validated playbooks instead of procedures improvised at every run. Most companies are still at stage one: prompt snippets. The compounding happens at stage three: packaged capability agents can run.
3/ Protocols: your handoffs, made machine-readable: what travels, from whom to whom, in what form, with which permissions.
And one requirement holds it all together: governance.
Shared AI infrastructure needs, in the authors' words, "operating-system-like governance": access control, oversight, conflict resolution. Guardrails and gates aren't compliance decoration. They're what makes the whole thing workable.
Here's why this matters to you more than any model announcement: your competitive advantage is in what you build around the model, not the model itself. Everything externalized — the memory, the skills, the protocols, the governance — is yours, and it's what will give you an edge. Or not.
A quick check: Is your decision history written anywhere your AI can read it? Are your ways of working codified in playbooks that all teams can leverage? Are agents running your low-value tasks and prepping the work for your teams?
This is the work I do with my clients and their teams: externalizing what makes them good — their memory, their skills, their handoffs — so their AI stops improvising and starts compounding.
If this resonates, let's talk!