langchain-ai / langchain-ai/deepagentsjs
Memory leak with dynamic sub agent middleware
- Dominant language
- TypeScript
- Stars
- 1.6k
- Forks
- 272
- Avg merge
- 1d 13h
- Merged PRs (30d)
- 38
Description
Hi,
I've implemented the dymanic middleware in my harness earlier, at https://github.com/avoidwork/madz/blob/main/src/agent/deepAgents.js#L249
In the past 2 days, with DeepSeek-V4-Flash running in madz, I've had crippling memory leaks - it looks like a more capable LLM has found a significant issue in the lib/middleware - it looks like it just keeps spawning - I haven't done a proper RCA yet though; I just woke up + got a coffee + found this:
Memory usage was flat before I added the dynamic middleware with DeepSeek-V4-Flash. I also want to point out I don't see this memory leak manifest with other LLMs like the Qwen family - I think DeepSeek-V4-Flash is simply doing something "better" which has found an edge case in the lib.
Contributor guide
Research direction
Start by reproducing the memory growth with the dynamic middleware and DeepSeek-V4-Flash, then compare it with the Qwen behavior described in the issue. Inspect lib/middleware alongside the linked harness entry point at src/agent/deepAgents.js#L249. Done means identifying the repeated spawning or retention path and confirming that memory remains stable after the fix.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- typescript
- Domain
- ai
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Active
- Clarity
- Needs clarification
- Newbie friendliness
- 35/100