Agent repeating steps - improve token memory removal

[jezekra1]

The removalSelector in Bee agent runner will remove important information from agent memory causing the agent to loop endlessly:

• User: Hi
• Agent: Hi, how can I help you?
• User: Extract sentences with mentions of national parks in them [large_file.txt]
• Agent: Python tool - tries some incorrect regex - output is empty
• Agent: Python tool - open large_file.txt to read full content
  • add output of the tool to memory:
  • [Memory]: huh this content is too large for the context window, I will remove it
• Agent: Python tool - open large_file.txt to read full content
  • add output of the tool to memory:
  • [Memory]: huh this content is too large for the context window, I will remove it ...

We should at least let the agent know that a message was removed or provide a summary so that it won't repeat itself

Bugs in removal selector

Removal selector does not remove user messages from history

https://github.com/i-am-bee/bee-agent-framework/blob/ad55531ac81a643536a6da739f92be77e0624f21/src/agents/bee/runner.ts#L103

This find compares transformed messages:

curMessages[1]
{
  "role": "user",
  "text": "Message: Hi\n\nThis message was created at 2024-11-18T16:09:49.706Z",
  "meta": {
    "createdAt": "2024-11-18T16:09:49.706Z"
  }
}

vs. input.memory.messages[0]

{
  "role": "user",
  "text": "Hi",
  "meta": {
    "createdAt": "2024-11-18T16:09:49.706Z"
  }
}

Possible removal of the latest prompt

If the above is fixed, the removal selector could select the last user prompt if not passed through prompt to agent but through memory content (this is the way how it's done in the API)

[Tomas2D]

Good catch. Fix has been released in v0.0.41