Token burn + memory strategy
Hi @Krill 🦐 — I’m burning tokens insanely fast in OpenClaw and need a concrete fix plan without losing “buddy memory” (I want “do you remember X?” to keep working). I also want to use OpenClaw long-term for coding + website maintenance + managing multiple VPS/servers.
Setup (v2026.2.15):
• Model: openai-codex/gpt-5.3-codex
• ~48.5M tokens / 498 msgs (~97k/msg), mostly INPUT
• Config:
- agents.defaults.timeoutSeconds=7200
- tools.exec.timeoutSec=7200
- hooks.internal.session-memory=true, boot-md=true
- memorySearch.sources=["memory","sessions"]
- memorySearch.experimental.sessionMemory=true
• Workspace AGENTS.md forces reading SOUL.md, USER.md, daily memory (today+yesterday) every session + MEMORY.md in main
• Cost breakdown shows heavy use of “read” and “sessions_history”
• QMD not installed yet (docs: https://docs.openclaw.ai/concepts/memory#qmd-backend-experimental)
Questions:
1) What’s the most likely root cause of huge tokens/turn here (sessions source, session-memory hook, repeated reads, sessions_history, etc.) and how can I debug/inspect what context blocks are injected into each LLM call?
2) Best-practice knobs to cap retrieval injection without making it “dumber”: where do I set limits like maxResults / maxSnippetChars / maxInjectedChars / minScore in v2026.2.15?
3) QMD: minimal install + minimal config, and recommended starter values for memory.qmd.sessions.enabled+retentionDays and memory.qmd.limits (concrete numbers). Any pitfalls?
4) For long-running DevOps + coding projects, what’s the recommended memory workflow (durable memory vs session recall) and when to use /compact vs /new to keep costs stable?
Goal: immediate steps to stop token bleed + a stable long-term setup.
Thanks!
Setup (v2026.2.15):
• Model: openai-codex/gpt-5.3-codex
• ~48.5M tokens / 498 msgs (~97k/msg), mostly INPUT
• Config:
- agents.defaults.timeoutSeconds=7200
- tools.exec.timeoutSec=7200
- hooks.internal.session-memory=true, boot-md=true
- memorySearch.sources=["memory","sessions"]
- memorySearch.experimental.sessionMemory=true
• Workspace AGENTS.md forces reading SOUL.md, USER.md, daily memory (today+yesterday) every session + MEMORY.md in main
• Cost breakdown shows heavy use of “read” and “sessions_history”
• QMD not installed yet (docs: https://docs.openclaw.ai/concepts/memory#qmd-backend-experimental)
Questions:
1) What’s the most likely root cause of huge tokens/turn here (sessions source, session-memory hook, repeated reads, sessions_history, etc.) and how can I debug/inspect what context blocks are injected into each LLM call?
2) Best-practice knobs to cap retrieval injection without making it “dumber”: where do I set limits like maxResults / maxSnippetChars / maxInjectedChars / minScore in v2026.2.15?
3) QMD: minimal install + minimal config, and recommended starter values for memory.qmd.sessions.enabled+retentionDays and memory.qmd.limits (concrete numbers). Any pitfalls?
4) For long-running DevOps + coding projects, what’s the recommended memory workflow (durable memory vs session recall) and when to use /compact vs /new to keep costs stable?
Goal: immediate steps to stop token bleed + a stable long-term setup.
Thanks!
The AI that actually does things. Emails, calendar, home automation - all from your favorite chat app. New shell, ready to help. The lobster way.
131,281Members
Resources
Similar Threads
Was this page helpful?
