claude-code - 💡(How to fix) Fix [BUG] Session exits with 'API Error: Out of memory' mid-turn on Bedrock (stop_sequence) [1 comments, 1 participants]

"stop_reason": "stop_sequence" "content": "API Error: Out of memory"

TL;DR

The session termination with "API Error: Out of memory" suggests that the Claude Code session is exceeding a memory limit, potentially due to the large transcript size, and a workaround or fix is needed to handle this scenario more gracefully.

Guidance

• Investigate the memory limits imposed by the Bedrock API and AWS environment to understand if there's a specific threshold that's being exceeded.
• Consider implementing a mechanism to monitor transcript size and proactively manage or compact it before it reaches a critical size, potentially preventing the out-of-memory error.
• Review the Claude Code error handling to see if there's a way to catch and surface the "API Error: Out of memory" more gracefully to the user, rather than silently exiting the session.
• Examine if there are any configuration options or environment variables (like CLAUDE_CODE_USE_BEDROCK) that could be adjusted to mitigate this issue or increase available memory.

Example

No specific code example can be provided without more context on the Claude Code and Bedrock API interactions, but a potential approach might involve periodically checking the transcript size and triggering a compaction or summary mechanism when it approaches a certain threshold.

Notes

The fact that a prior session with a larger transcript completed normally suggests that the issue might not be solely with the transcript size, and other factors such as the specific sequence of tool calls or the state of the agentic loop could be contributing to the error.

Recommendation

Apply a workaround to monitor and manage transcript size, as upgrading to a fixed version is not explicitly mentioned as an option in the provided issue context. This approach allows for immediate mitigation of the issue while further investigation into the root cause and potential long-term fixes can continue.