122 lines
2.8 KiB
Markdown
122 lines
2.8 KiB
Markdown
# How to Never Hit OpenClaw's 150K Context Limit
|
|
|
|
**Source:** X Thread by [Johann Sathianathen](https://x.com/johann_sath)
|
|
**Date:** February 21, 2026
|
|
**URL:** https://x.com/johann_sath/status/2025440759416045702
|
|
|
|
---
|
|
|
|
## The Problem
|
|
|
|
Most OpenClaw users hit 150K context in a single conversation and wonder why their agent gets dumber. The solution isn't a bigger window—it's never needing one.
|
|
|
|
---
|
|
|
|
## The 5 Strategies
|
|
|
|
### 1. Add "Orchestrator" Pattern to SOUL.md
|
|
|
|
**File:** `SOUL.md`
|
|
**Add:** "You are the orchestrator. Subagents execute."
|
|
|
|
**Why it works:**
|
|
- Main session stays lean
|
|
- Heavy work runs in fresh context windows
|
|
- Subagents spawn, work, report back, context dies
|
|
|
|
**Result:** Main session runs 24/7 without bloat
|
|
|
|
---
|
|
|
|
### 2. Use BRAIN.md as External Memory
|
|
|
|
**File:** `BRAIN.md`
|
|
**Pattern:** Agent reads & writes instead of remembering everything in-context
|
|
|
|
**Why it works:**
|
|
- Offloads memory from context to disk
|
|
- Agent references files instead of recalling
|
|
- Structured storage for complex state
|
|
|
|
---
|
|
|
|
### 3. Set Up Heartbeats
|
|
|
|
**Pattern:** Fast check-ins (<3s) that don't load files unless you're idle
|
|
|
|
**Implementation:**
|
|
- Only check critical items
|
|
- Skip file reads unless necessary
|
|
- Batch multiple checks together
|
|
|
|
**Benefit:** Keeps session responsive without context bloat
|
|
|
|
---
|
|
|
|
### 4. Run Cron Jobs Isolated
|
|
|
|
**Pattern:** Each cron job gets its own session
|
|
|
|
**Why it works:**
|
|
- Never bloats main thread
|
|
- Independent execution contexts
|
|
- Parallel processing without interference
|
|
|
|
---
|
|
|
|
### 5. Delegate Everything to Subagents
|
|
|
|
**Pattern:** Subagent spawns → does work → reports back → context dies
|
|
|
|
**Flow:**
|
|
1. Main session identifies task
|
|
2. Spawns subagent with specific goal
|
|
3. Subagent completes work in isolation
|
|
4. Reports results back
|
|
5. Subagent context garbage collected
|
|
|
|
---
|
|
|
|
## The Result
|
|
|
|
Johann's main session:
|
|
- Runs 24/7
|
|
- **Rarely passes 30K context**
|
|
- Stays responsive and smart
|
|
|
|
**Key insight:** The trick isn't a bigger window. It's never needing one.
|
|
|
|
---
|
|
|
|
## Implementation Checklist
|
|
|
|
- [ ] Add orchestrator pattern to SOUL.md
|
|
- [ ] Create/improve BRAIN.md for external memory
|
|
- [ ] Audit heartbeats for efficiency (<3s checks)
|
|
- [ ] Review cron jobs for isolation
|
|
- [ ] Identify opportunities for subagent delegation
|
|
|
|
---
|
|
|
|
## Related Resources
|
|
|
|
- **SOUL.md Guide:** https://soul.md (defining AI identity)
|
|
- **Johann's Weekly Newsletter:** https://johann.fyi
|
|
- **Author:** Building dmpro.ai (ex-Cisco), 1x exit, founder @jars_global
|
|
|
|
---
|
|
|
|
## Key Takeaway
|
|
|
|
Context management isn't about having more—it's about using less by:
|
|
1. Orchestrating instead of executing
|
|
2. Externalizing memory
|
|
3. Efficient polling
|
|
4. Isolated scheduling
|
|
5. Delegating to subagents
|
|
|
|
---
|
|
|
|
*Compiled from X thread by Johann Sathianathen*
|
|
*Engagement: 5 replies, 5 reposts, 60 likes, 199 bookmarks, 5,198 views*
|