Question about web benchmark implementation and tool integration in AtomMem

[GenSouKa1]

Hi, thanks for the great work!

I’ve been reading both the paper and the codebase carefully, and I have a question regarding the web benchmark part (Asearcher / GAIA / WebWalkerQA).

From the paper, it seems that:

• AtomMem is evaluated on multi-turn web tasks
• The agent is equipped with tools like search engine and URL reader
• The environment allows up to 40 tool calls per task

However, in the current codebase, I can only find implementations related to long-context QA (e.g., document chunking, memory + retrieval pipeline). I wasn’t able to locate:

1. The implementation of the web environment wrapper
2. The definition of external tools (e.g., search / URL reader)
3. How different web datasets (Asearcher / GAIA / WebWalkerQA) are unified into a common interaction protocol
4. The rollout loop for multi-turn tool-augmented interaction

So I’m wondering:

• Is the web benchmark part not included in the current release?
• Or is it implemented in another repository / branch?
• If it exists, could you point me to the relevant modules?

Additionally, I’m particularly interested in how you handle:

• Tool abstraction (e.g., unified XML action → actual API call)
• Environment feedback format (observation construction after each tool call)
• Compatibility across different web benchmarks

Thanks again for the excellent paper — the idea of modeling memory as a decision process is very inspiring!

Best regards.