Prediction · On the road to AGI
Advanced AI systems will keep learning after they are deployed, with long-term memory as a standard part, instead of staying frozen after training.
Why we think so
Agents with memory and reflection already behave with continuity across time (generative agents). Retraining large models from scratch is slow and costly, so learning in deployment is the natural next step. We give this low confidence because doing it safely, privately and without forgetting is unsolved.
What would prove us wrong
Frozen models plus short context staying the norm for the most capable systems.
Review notes
21 October 2026: Memory carried across steps (state preservation and compaction) decided the ARC-AGI-3 result, a sign that memory is becoming central. True learning after deployment is not yet standard.
- Horizon
- Medium term. About 3 to 10 years out.
- Confidence
- Low confidence
- Status
- Open. Not yet decided either way.
- Made
- 21 October 2026
- Last reviewed
- 21 October 2026
- Research programs
- Memory and continual learning
The evidence it rests on
Public results from teams around the world. Read the sources and judge for yourself.
· Stanford University and Google researchers (arXiv)
Generative agents simulate a small society
Twenty-five agents with memory, reflection and planning lived in a simulated town and showed believable individual and group behaviour.
What it means for AGI: memory and reflection give agents continuity, and groups of agents produce behaviour no single prompt produces.
Source: arxiv.org(opens external site)
· ARC Prize Foundation (testing OpenAI's model)
GPT-6 Astra scores 99.9% on ARC-AGI-3, with a caveat
ARC Prize verified 62.7% with its standard harness and 99.9% with a harness that keeps the model's reasoning state between steps; in one setting it used fewer actions than the median person on 96% of levels. ARC Prize states it is not claiming this is AGI.
What it means for AGI: the gap on learning new tasks closed within six months of the test's launch. The jump from 62.7% to 99.9% came from how the system keeps memory, not from a new model, which supports the view that AGI is a system, not one model. The run cost about $18,800, far more than a person would.
Think we are wrong?
Send us the evidence. If it changes our view, we update the prediction and say why.