Continual learning
Repeat bounded model updates as reviewed data arrives, then evaluate each candidate before accepting it.
Continual learning updates a model over time using new data. In OpenPond, a Model's learning policy controls the source, supported training setup, schedule, spending and candidate decision. A new conversation or Dataset row does not change weights by itself.
The objects in one cycle
| Object | Role |
|---|---|
| Dataset | Reviewed, versioned examples or tasks used as inputs. |
| Experiment | One configured evaluation, training, modification or inspection, with retained results. |
| Modification | What that experiment does to the model, weights or activations. |
| Model | The saved identity and versions that subsequent work can use. |
| Learning policy | When supported updates run and how their candidates are evaluated and accepted. |
Where to configure it
- Open the Model's learning settings. Reuse its source, schedule, budget and approval controls; do not schedule the same update from several places.
- A Dataset can supply multiple Models. The learning policy identifies both the Dataset and target Model; a Dataset is not itself a trained model.
- Each cycle retains its own inputs, spending and results. Opening an experiment or saving a draft does not authorize another cycle.
- Keep holdout evaluation separate from training data. Check earlier skills as well as the new task before accepting a candidate.
Example
A support Model receives reviewed examples each week:
- Publish the new support examples in a Dataset.
- A supported learning policy starts a bounded training experiment.
- Compare its candidate with the current Model on separate support and regression tasks.
- Approve a passing candidate as the next saved version. Activate serving only under the policy's separate explicit authority.
The result is a version and evaluation record, not a promise that every update improves the model. Insufficient data, failed checks or exhausted credits can prevent a cycle from training or accepting a candidate.
Which methods can repeat?
Available scheduling depends on the saved training configuration and runtime. SFT and GRPO have their own policy and configuration gates. A method appearing in the modification chooser does not establish recurring-training support. For example, an exact Engram parent can support a new explicit experiment without implying that Engram is available to the scheduler.
For imported coding-agent conversations, follow Continual learning from harnesses. Recurring Harness review changes the agent's surrounding behavior; it does not itself update model weights.