facing high token burn challenges · TopOfMind AI Builders

facing high token burn challenges

Hey folks/quick question for anyone running Clawdbot.

what do you consider a healthy “always-on” context size per session from a token + cost perspective?

I’m currently tightening (or at least trying to/lol) things to:

~15K tokens for active context
Last 3–5 messages only
Aggressive summarization + compaction beyond that
Using a cheaper model for non-thinking tasks (summaries, formatting, validation)

Curious:

Where do you cap context in practice?
Do you rely on auto-compaction (maybe the gateway helps to compact but the bot adds all contexts which always pushes the size of contexts) or manual summaries?
Any gotchas you’ve hit with session memory blowing up costs?

Would love to hear real-world numbers vs theory.

8 comments

facing high token burn challenges