Subject: Context compaction firing prematurely despite high contextTokens setting
Environment:
Clawdbot version: 2026.1.24-3
Platform: Windows 10, Node v24.1.0
Channel: Discord
Model: claude-opus-4-5
Current Config (relevant settings):
{
"agents": {
"defaults": {
"contextTokens": 150000,
"memorySearch": { "enabled": true },
"compaction": {
"mode": "safeguard",
"memoryFlush": { "enabled": true }
}
}
}
}
Issue:
Context compaction fires after only 2-3 tool-heavy tasks (e.g., reading files, searching Discord channels), despite contextTokens set to 150,000. The agent loses conversation context mid-task and requires re-explanation of prior work.
Observed behavior:
Compaction triggers seemingly at random, not near the 150k limit
Token usage isn't approaching max before truncation occurs
Agent writes to memory files during tasks, but compaction still wipes active context
Feels like significantly less context than using Claude Code directly in terminal
Attempted fixes:
Increased contextTokens from default to 150,000
Enabled memoryFlush (to save context before compaction)
Enabled memorySearch for recovery
Agent writes progress to memory/YYYY-MM-DD.md files during work
Questions:
Is there a way to extend/prevent compaction entirely for longer sessions?
What actually triggers compaction โ is it token count, message count, or something else?
Are there additional config options to increase context retention?
Is there a way to see actual token usage before compaction fires?
Clawdbot version: 2026.1.24-3
Platform: Windows 10, Node v24.1.0
Channel: Discord
Model: claude-opus-4-5
Current Config (relevant settings):
{
"agents": {
"defaults": {
"contextTokens": 150000,
"memorySearch": { "enabled": true },
"compaction": {
"mode": "safeguard",
"memoryFlush": { "enabled": true }
}
}
}
}
Issue:
Context compaction fires after only 2-3 tool-heavy tasks (e.g., reading files, searching Discord channels), despite contextTokens set to 150,000. The agent loses conversation context mid-task and requires re-explanation of prior work.
Observed behavior:
Compaction triggers seemingly at random, not near the 150k limit
Token usage isn't approaching max before truncation occurs
Agent writes to memory files during tasks, but compaction still wipes active context
Feels like significantly less context than using Claude Code directly in terminal
Attempted fixes:
Increased contextTokens from default to 150,000
Enabled memoryFlush (to save context before compaction)
Enabled memorySearch for recovery
Agent writes progress to memory/YYYY-MM-DD.md files during work
Questions:
Is there a way to extend/prevent compaction entirely for longer sessions?
What actually triggers compaction โ is it token count, message count, or something else?
Are there additional config options to increase context retention?
Is there a way to see actual token usage before compaction fires?
The AI that actually does things. Emails, calendar, home automation - all from your favorite chat app. New shell, ready to help. The lobster way.
83,815Members
Resources
Similar Threads
Was this page helpful?
