Fix Context Bloat After Onboarding: 8 Tips From the Community
A recent discussion in the OpenClaw Discord revealed a common pain point: users running through onboarding, then watching their context window balloon to 130k+ tokens within just a few messages. Here's what the community shared about fixing it.
The Problem
User SchaduwHond described the issue: "132k tokens in, in 3 messages鈥攖hat is insane." Despite having a tiny memory.md file (just 5 lines with basic info), the token usage was burning through their API credits fast.
The root cause? openclaw onboard sets up a generic baseline, but it enables a lot by default. Personalization means optimizing for your workflow, not running everything at once.
Community Solutions
1. Debug Your Context First
Run /context to see exactly what's being loaded into each prompt. This shows you what's eating tokens before you start cutting.
2. Disable Tools and Skills You Don't Use
As reddev put it: "It's very tempting to enable all the tools/skills." But every enabled tool adds to your system prompt. Be ruthless鈥攊f you're not using home automation today, disable that skill.
3. Use Subagents With Cheaper Models
Don't run Opus for everything. Spawn subagents on Sonnet or Flash for routine tasks. The main agent orchestrates; the subagents do the heavy lifting on cheaper models.
4. Optimize Heartbeat Model
Your heartbeat checks don't need a frontier model. Configure a cheaper model for heartbeat polling to save significant costs over time.
5. Create Documentation Outside the System Prompt
Casimir1904 shared a key insight: "You can create documentation that isn't loaded into the system prompt." Point your agent to reference docs it can read when needed, rather than injecting everything upfront.
6. Load Daily Memory Instead of memory.md
"You can tell the AGENTS.MD to load the daily memory not the memory.md but it can use memory_search in that memory.md."
This is powerful: your daily notes are lightweight and contextual, while memory.md becomes a searchable archive accessed only when needed via semantic search.
7. Set Up Model Routing
Configure different models for different task types. Simple queries get cheap models; complex reasoning gets premium models. This requires some setup but pays dividends.
8. Ask Your Agent
Seriously鈥攜our agent knows OpenClaw's internals. As reddev advised: "Ask your bot what's possible." Point it at the docs and let it suggest optimizations for your specific setup.
The Mindset Shift
The thread surfaced an important truth: openclaw onboard is a starting point, not the finish line. The users who get the most value are the ones who invest time tuning their setup.
As Ineffigy noted: "Always a good idea to document all your customizations so you can re-create them if need be or to better see all that has been done."
Thanks to SchaduwHond, reddev, Casimir1904, and Ineffigy for the discussion that inspired this post.
Comments (0)
No comments yet. Be the first to comment!