Memory Leak

[Michaelzag]

App Version

3.16.6

API Provider

Anthropic

Model Used

Gemini 2.5 Pro -> Claude 3.7 Thinking

🔁 Steps to Reproduce

I've been using orchestrator mode and swapping to an Anthropic 3.7 with thinking mode enabled.
Normally I'm running with 500k context or so. Maybe it's been chopped once or twice with sliding window.
When a sub task is completed, sometimes it doesn't properly swap back to Orchestrator mode, it thinks it does, and tries to return to the 500k+ context window, but in the code/claude 3.7 context window.
VS code freezes and begins eating all my remaining ram. This windows pc has 96 gb of ram and I'm usually using 30-40gb.
I cant reproduce it, but it's happens about once every 2 hours.
It has happened after an extended session and it has happened at the beginning-ish of a session. before the first gemini context chop.
I'll post more details if I can next time it happens.

💥 Outcome Summary (Optional)

No response

📄 Relevant Logs or Errors

[dosubot]

Hello @Michaelzag! I'm Dosu, your friendly bot here to help with bugs, answer questions, and support you while you wait for a human maintainer. Let’s tackle this issue together!

It seems like you're experiencing a memory leak issue when swapping between orchestrator mode and the Anthropic model, particularly when handling large context windows. Unfortunately, I couldn't find specific documentation on memory leak issues related to large context windows in orchestrator mode or error handling procedures when a mode swap fails. However, I can suggest a few general steps to help troubleshoot and potentially mitigate the issue:

1. Context Window Management: Ensure that the context window size is properly managed and adjusted. The default context window size might be too small for your needs, and you may need to configure it to a larger size, such as 32k, using the command /set parameter num_ctx 32768 ^[1].

2. Monitor Resource Usage: Use system monitoring tools to track memory usage over time. This might help identify patterns or specific actions that lead to increased memory consumption.

3. Logs and Debugging: Enable detailed logging if possible, to capture more information about the state of the application when the issue occurs. This might provide insights into what triggers the memory leak.

4. Check for Updates: Ensure that you are using the latest version of the software, as updates may contain fixes for known issues.

5. Report Details: If you can reproduce the issue or gather more details, consider submitting a detailed bug report to help the maintainers identify and address the problem.

If you have any more specific questions or need further assistance, feel free to ask!

To continue talking to Dosu, mention @dosu.

---

^{Help Dosu learn! Give it feedback: Great Response | Irrelevant Answer | Incorrect Sources | Too Verbose | Hallucination | Bug Report | Other}

[Michaelzag]

While I didnt replicate it, I think I'm close.
I was in Orchestrator mode, and pasting in console logging errors.
It was dispatching sub tasks and that was a bit tedious so I swapped to Debug mode.
Gemini Orchestrator -> Gemini Debug Mode
Compared to the Claude days of 200k context, I am really abusing the chat window and pasting in really long error logs...
Like 400k tokens into the window.

The specific interface of debug mode was getting annoying. I'm pasting in the console logs because I dont want to read them, Debug mode was asking me specific questions about that 400k context I just pasted in.
Instead of answering the quesiton debug mode asked. I swapped to code mode and was getting ready to tell it to just fix it.
Code mode is still tagged to Claude 3.7

When I sent the next message it froze and has lost it's context. It didnt eat all my memory though.

[Fastidius-au]

FYI i use MCP servers alot MemoryBank custommodes and lots of prompting and such.

I think my issue might be a leek too and seems to be in code mode with vscode current.13 may. i say code mode because I use architect and orchestra but it only happened to me in Code mode and it was not on mode change....suspect it may be gemini end since it been a shitfight the last few days with 503 502 not available etc but it caught my eye when I looked at bugs

it calls an API then dies......context can be anywhere between 400k and 1.2 mill......i blew the cap a few times and I have had it at qroung 600-800K fillup stage.

I'm not the original poster nor did I think it was roo because I have symlinks in my codebase and figured it sometimes died when I went to a network area and had spun down or something but I have been having it consistently and it happened a few times on the first couple of words of the stream....Hope that help

[Michaelzag]

Humm, did it again.
After a boomerang task failed to return, I went back to the original chat and told it that the task failed and to fix a few things and verify it.
Then it dispatched a subtask to update some docs and verify the changes were completed by the previous subtask.
When the second task was completing it ran out of memory. This Orchestrator wasn't overloaded.