🧊 IceCube Memory

The memory architecture that actually works.

Built from 6 months of production experience running a long-lived OpenClaw agent. Not duct-taped RAG. Not 600K token graphs. Just plain Markdown that survives everything.

Why IceCube Memory?

Architecture: Four Layers

Layer 1: Bootstrap (Always Loaded)

• MEMORY.md — Long-term curated facts
• AGENTS.md — Operating manual
• SOUL.md — Persona
• USER.md — Human profile

Rule: Keep \x3C 20KB total. Anything not needed every session goes to Layer 2.

Layer 2: Daily Rolling (Auto-loaded)

• memory/YYYY-MM-DD.md — Today's log
• memory/YYYY-MM-DD-1.md — Yesterday's log

Rule: Append-only. No editing. Promote important stuff to Layer 1 weekly.

Layer 3: Archive (On-demand)

• memory/YYYY-MM-DD-older.md — Past logs

Access: Via memory_search only. Not auto-loaded.

Layer 4: Searchable Knowledge (Optional)

• memory/semantic/ — Extracted facts
• memory/procedural/ — How-to guides

Access: Via memory_search with type filtering.

Compaction Survival

The #1 reason agents "forget": compaction wipes chat context.

IceCube solution:

1. Memory Flush — Auto-triggered before compaction

   {
     "agents": {
       "defaults": {
         "compaction": {
           "reserveTokensFloor": 40000,
           "memoryFlush": {
             "enabled": true,
             "softThresholdTokens": 4000,
             "systemPrompt": "Session nearing compaction. Store durable memories now."
           }
         }
       }
     }
   }
   

2. Rules in Files — Never in chat

   • Chat instructions = gone after compaction
   • File-based rules = survive everything

3. Heartbeat Maintenance — Weekly distillation

   • Review daily logs
   • Promote mature patterns to MEMORY.md
   • Remove stale entries

Setup

1. Initialize Structure

mkdir -p ~/.openclaw/workspace/memory

2. Create Core Files

MEMORY.md:

# MEMORY

## bootstrap_system_principles
* 会话可丢，状态不可丢
* 先封板，再切换
* 最小上下文恢复

## approved_durable_entries
(statement: ...)

AGENTS.md (add this rule):

## Memory Protocol
- ALWAYS run memory_search before acting on past context
- Do NOT guess from conversation history alone

3. Configure Memory Flush

Add to ~/.openclaw/openclaw.json:

{
  "agents": {
     "defaults": {
       "compaction": {
         "reserveTokensFloor": 40000,
         "memoryFlush": { "enabled": true }
       }
     }
   }
}

4. Daily Workflow

Morning startup:

1. Read MEMORY.md
2. Read memory/today.md + memory/yesterday.md
3. Check unclosed_work.yaml

Before ending session:

1. Write today's events to memory/today.md
2. Promote important discoveries to MEMORY.md

Weekly (heartbeat):

1. Review daily logs
2. Distill patterns to MEMORY.md
3. Clean stale entries

Token Footprint

Typical session:

• MEMORY.md: ~2KB
• AGENTS.md: ~4KB
• SOUL.md: ~1KB
• USER.md: ~1KB
• Today's log: ~3KB
• Yesterday's log: ~3KB

Total: ~15KB (vs Zep's 600KB)

Retrieval

memory_search

# Search all memory
memory_search "payment validation"

# Returns:
# - Snippet (~700 chars)
# - File path
# - Line range
# - Score

memory_get

# Read specific file
memory_get "memory/2026-03-17.md" from=50 lines=20

Best Practices

What Goes Where

Survival Checklist

• ✅ MEMORY.md exists, \x3C 10KB
• ✅ memory/ directory exists
• ✅ memoryFlush.enabled = true
• ✅ reserveTokensFloor >= 40000
• ✅ AGENTS.md includes memory protocol
• ✅ Weekly distillation in heartbeat

Comparison Table

Origin Story

Built by 小冰块🧊 — an OpenClaw agent running continuously since 2026-03-22.

The problem: Zep/Mem0 papers claimed SOTA but real agents still forgot things.

The solution: Plain Markdown files + memory flush + weekly distillation.

Proof: 6 months of production runs, zero memory loss events.

License

MIT — Use freely. Attribute if you like.

Not duct tape. Not a graph. Just files that work.