Question about the paper and the code

[Yangseung]

I understood that the RL‑tuned model determines an action and then executes that action (e.g., delete, update, etc.).
However, from the code, it seems like the model only learns to choose the action, not to perform it. Is that correct?

Also, in memory_server.py, the _get_analysis_prompt function looks like part of the memory extraction process.
Is this step handled by the RL‑tuned model, or is it done by a larger base model?

[Yangseung]

Also, we found the error when using the script python main.py --agent_config config/memalpha-qwen3-4b_agent_0.05-0.1.yaml --dataset memalpha

There is not tokenizer in memory_agent_template when using the fine-tuned RL model

[Yangseung]

• I understand that Mem-alpha creates new memory by prompting with prior core, semantic, and episodic memories to process the current instance. If the memory becomes very large, we may be unable to include all three types; conversely, if we keep including everything, the prompt will contain a lot of irrelevant information, which could degrade accuracy.

[Yangseung]

Also, I have one additional question.
Why did not use the official MemoryAgentBench for evaluation?
I think Mem-alpha is also validated to the official dataset.
Is that any reason?

[wangyu-ustc]

Hi for the first question:

I understood that the RL‑tuned model determines an action and then executes that action (e.g., delete, update, etc.).
However, from the code, it seems like the model only learns to choose the action, not to perform it. Is that correct?

Also, in memory_server.py, the _get_analysis_prompt function looks like part of the memory extraction process.
Is this step handled by the RL‑tuned model, or is it done by a larger base model?

Yes the model only chooses the action. For all the agents like Claude-Code or Cursor, the model is always choosing the action and there is the environment that is perfoming it, this is the standard way to adopt LLM agents.
"_get_analysis_prompt" is part of the process of getting the reward $$r_4$$ and this process is done by a larger model (Qwen3-32B).

[wangyu-ustc]

Also, we found the error when using the script python main.py --agent_config config/memalpha-qwen3-4b_agent_0.05-0.1.yaml --dataset memalpha

There is not tokenizer in memory_agent_template when using the fine-tuned RL model

Sorry let me fix this issue in the following days.

[wangyu-ustc]

I understand that Mem-alpha creates new memory by prompting with prior core, semantic, and episodic memories to process the current instance. If the memory becomes very large, we may be unable to include all three types; conversely, if we keep including everything, the prompt will contain a lot of irrelevant information, which could degrade accuracy.

In our code we do not always let the model see all the memories, which may defeat the original goal of our paper. You can see the code here: https://github.com/wangyu-ustc/Mem-alpha/blob/main/main.py#L49
We only allow the model to see the most recent 10 memories when seeing the new chunk.

[wangyu-ustc]

Also, I have one additional question.
Why did not use the official MemoryAgentBench for evaluation?
I think Mem-alpha is also validated to the official dataset.
Is that any reason?

We did use the offical dataset. We only excluded the category Conflict Resolution as it remains synthetic.

[wangyu-ustc]

Hi, as for this issue: #3 (comment)
We have fixed it in the newest commit, please check it out, thanks!