[Question]: Any plans to support Acontext as a data source for Agent RL training?

[benwuhua]

First of all, thank you for building such a great tool for context and interaction management.

I would like to ask whether there are any plans to support Acontext as a data source for Agent Reinforcement Learning (Agent RL) training.

In many emerging agent ecosystems, teams are exploring how to leverage structured and semi-structured interaction data, context traces, and user behavior logs for RL-based training (e.g., RLHF/RLAIF/RLH and other agent-centric RL frameworks). If Acontext could eventually provide:

Data export interfaces (e.g., conversations, action sequences, event logs)

Optional schemas or standardized formats suitable for RL training

Access control and sampling mechanisms

Potential support for streaming or incremental data access

…this would significantly expand its usability for agent training and evaluation workflows.

I’m wondering whether such capabilities are on your roadmap or if you have any initial thoughts about the feasibility or direction.

[gusye1234]

Thanks for this great question！

Data export interfaces (e.g., conversations, action sequences, event logs)

Yes, it's definitely on our roadmap, just not in the next two minor versions(v0.0 and v0.1).

The data exporter will support export the context data with task annotations, which I also believe Acontext will provide a great form of data for RL training.

Optional schemas or standardized formats suitable for RL training

I'm not an expert in RL, would you tell me more about the where you want the schema to apply?

Access control and sampling mechanisms

We're on it!

Potential support for streaming or incremental data access

I think Acontext is already incremental? Every session is independent to each other.