[BUG] Working memory leaks into text stream

1 Reply
souvikinator
souvikinatorOP13h ago
TLDR; Hard to reproduce cuz it's happening very random times but I can share traces from langfuse if you want. here is an excerpt:
[redacted], huh. nice, small world.
what’s your exact role there — title and what you actually do day to day?{
"memory": "# User Profile\n\n## Personal Information\n- Name: Souvik\n- Email:\n- Phone:\n- Location (City/Country):\n- Current Timezone:\n- Preferred Language:\n- LinkedIn URL:\n- GitHub URL:\n\n## Professional Background\n- Current Role:\n- Current Company: [redacted]\n- Years of Experience:\n- Notice Period:\n- Key Skills and Technologies:\n- Notable Projects or...[whole working memory]
[redacted], huh. nice, small world.
what’s your exact role there — title and what you actually do day to day?
[redacted], huh. nice, small world.
what’s your exact role there — title and what you actually do day to day?{
"memory": "# User Profile\n\n## Personal Information\n- Name: Souvik\n- Email:\n- Phone:\n- Location (City/Country):\n- Current Timezone:\n- Preferred Language:\n- LinkedIn URL:\n- GitHub URL:\n\n## Professional Background\n- Current Role:\n- Current Company: [redacted]\n- Years of Experience:\n- Notice Period:\n- Key Skills and Technologies:\n- Notable Projects or...[whole working memory]
[redacted], huh. nice, small world.
what’s your exact role there — title and what you actually do day to day?
agent literally give this as output, and once this is done it again outputs. Notice how it repeats the same. This only happens when the agent starts processing/generating text and then makes tool call. It doesn't happen when agent makes tool call first.

Did you find this page helpful?