dvcrn OpenClaw Memory Resilience
Covers the two highest-impact layers of memory durability for OpenClaw agents.
The Four Memory Layers (quick reference)
| Layer | What it is | Survives compaction? |
|---|---|---|
| Bootstrap files (SOUL.md, AGENTS.md, etc.) | Injected from disk at every session start | Yes — reloaded from disk |
| Session transcript (JSONL on disk) | Conversation history | Lossy — summarized on compaction |
| LLM context window | What the model currently sees | No — fixed size, overflows |
| Retrieval index (memory_search / QMD) | Searchable index over memory files | Yes — rebuilt from files |
Core rule: If it's not written to a file, it doesn't exist after compaction.
Three Failure Modes
- Failure A — Instruction only existed in conversation, never written to a file. Most common.
- Failure B — Compaction summarized it away. Lossy. Nuance lost.
- Failure C — Session pruning trimmed old tool results. Temporary, lossless.
Diagnostic: Forgot a preference? → Failure A. Forgot what a tool returned? → Failure C. Forgot the whole thread? → Failure B.
Layer 1: Gateway Compaction Config
Apply via gateway config.patch. This is a global default — applies to all agents:
{
"agents": {
"defaults": {
"compaction": {
"mode": "safeguard",
"reserveTokensFloor": 40000,
"memoryFlush": {
"enabled": true,
"softThresholdTokens": 4000,
"systemPrompt": "Session nearing compaction. Store durable memories now.",
"prompt": "Write any lasting notes to memory/YYYY-MM-DD.md; reply with NO_REPLY if nothing to store."
}
},
"contextPruning": {
"mode": "cache-ttl",
"ttl": "1h"
}
}
}
}
See references/config-explained.md for why each value is set the way it is.
Layer 3: File Architecture
See references/file-architecture.md for the full pattern — bootstrap files, daily notes, archival, and QMD indexing.
Context Footer Standard
Add to every agent's SOUL.md. This gives you real-time visibility into context fill and compaction count — the single fastest way to know when to act before you hit the bad compaction path.
## Context Management (Auto-Applied)
**Every response:** fetch live status via `session_status`, append footer: `🧠 [used]/[total] ([%]) | 🧹 [compactions]`
- Auto-clear: **85% context** OR **6 compactions**
- Warn: **70% context** OR **4 compactions**
- Before clearing: file critical info to memory, then reset
See references/context-footer.md for why this matters and how to tune the thresholds.
Diagnosing Problems
Run /context list in any OpenClaw session to see:
- Which bootstrap files loaded and at what size
- Whether any files are TRUNCATED (over per-file limit)
- Total injected chars vs raw chars
If a file shows TRUNCATED: reduce it or raise bootstrapMaxChars in config.
If a file is missing entirely: check the workspace path in agent config.