lost vine
Hi everyone,
Has anyone experienced a sudden session reset?
The error message:
"Session auto-reset — the conversation exceeded the maximum context size and could not be compressed further. Your next message will start a fresh session."
Details:
Reset happens out of nowhere
Conversation context is lost
Can occur at any time (no specific pattern)
If you've encountered this, please share:
What platform/LLM are you using?
Did you find a fix?
Thanks! 🙏
junior bridge
what is your ctx length set too?
lost vine
adding error message
2026-05-11 18:11:59,421 INFO gateway.run: response ready: platform=discord chat=1503327091377307659 time=14.2s api_calls=2 response=69 chars
2026-05-11 18:11:59,422 INFO gateway.run: Skipping transcript persistence for context-overflow failure in session 20260511_180322_79e3bd92 to prevent session growth loop.
2026-05-11 18:11:59,423 INFO gateway.run: Auto-resetting session 20260511_180322_79e3bd92 after compression exhaustion.
2026-05-11 18:11:59,442 INFO gateway.platforms.base: [Discord] Sending response (227 chars) to 1503327091377307659```@junior bridge
Environment:
OS: Linux (Ubuntu/Debian)
Installation: Docker on VPS
Provider/Model: MiniMax-M2.7 via minimax
Platform: Self-hosted Hermes Agent on VPS
What I was trying to do:
Use Hermes Agent for ongoing conversations and task execution. After several message exchanges or complex tasks, the session appears to reset or lose context mid-conversation.
What happened instead:
When context length approaches the model's limit (or during long multi-step tasks), the session resets — conversation history is lost and the agent appears to start fresh, as if the context window was exceeded and the session was restarted.
Steps to reproduce:
Engage in a conversation with multiple back-and-forth exchanges (10+ turns)
Or run a complex multi-step task that generates significant tool outputs
At some point, session resets with no warning
What I already tried:
Restarting the gateway (hermes gateway restart)
Using shorter, more focused prompts
Running hermes debug share to gather logs
Relevant logs:
Debug Report: https://paste.rs/4dMyj
agent.log: https://dpaste.com/ATZQQRJKF
gateway.log: https://paste.rs/FGl9E
junior bridge
@past rock what happened here ?
past rockBOT
📄 web_extract: "https://paste.rs/4dMyj"
past rockBOT
<@1498667414873968700> what happened here ?
logs tell a pretty clear story. here's what happened:
Root cause: compression.threshold: 0.9 is too aggressive
MiniMax-M2.7 has a ~128k context window. Your sessions keep hitting ~105-110k tokens, at which point Hermes tries to compress the transcript. With compression.threshold: 0.9, it waits until the context is 90% full before triggering compression — that's cutting it extremely close. By the time it tries to compress, there's no room left to do the work, so it fails and resets.
Contributing factors from your logs:
agent.max_turns: 9999 means sessions run indefinitely, accumulating massive contextThe fix:
Lower your compression threshold so it kicks in earlier:
hermes config set compression.threshold 0.7
Or even 0.6 given your setup. This gives the compressor room to actually do its job before you're jammed against the ceiling.
Also consider:
hermes config set agent.max_turns 50
9999 turns is effectively unlimited — long-running tasks are the direct cause of 100k+ token sessions.
The kanban duplicate column max_retries error is a separate known migration bug — it's benign (won't crash anything) but worth noting.
The gateway restart chaos was caused by you repeatedly trying to restart while an agent was mid-run. Each drain hit the 60s timeout with an active agent, requiring SIGKILL. Not a bug, just rough to deal with when sessions are hanging due to context issues.
tl;dr — drop compression.threshold to 0.65-0.7 and set a sane max_turns limit. The resets should stop.
junior bridge
Let us know if this helps !
lost vine
halo @junior bridge @past rock thx for the solution, it works!!