OpenClaw Memory

Session-first memory curator for OpenClaw. Keeps RAM clean, recall precise, and durable knowledge safe.

Security Scan

The skill's instructions, requirements, and behavior are internally consistent: it is an instruction-only memory curator that instructs the agent to persist curated notes to disk and use platform memory search APIs, and it does not request unrelated credentials or installs.

Purpose & Capability

Name and description match the SKILL.md: the skill is a session-first memory curator whose declared behavior is to extract durable items from session memory and write them to disk (MEMORY.md and daily logs). It asks only for enabling session memory in agent config, which is coherent with its purpose.

Instruction Scope

SKILL.md stays on-topic: it describes when/how to write durable memory, retrieval via memory_search/memory_get, and a pre-compaction flush routine. It instructs the agent to append files on disk and to modify an agent config flag (clawdbot config set) — both are consistent with a memory curator. It explicitly forbids storing secrets and copying raw chat into memory. There is no guidance to read arbitrary system files or exfiltrate data to external endpoints.

Install Mechanism

Instruction-only skill with no install spec and no code files. Nothing is downloaded or written by an installer, which minimizes risk and is appropriate for the stated functionality.

Credentials

The skill requires no environment variables, credentials, or external service keys. The SKILL.md also contains explicit privacy rules (do not store secrets), so requested access is minimal and proportionate.

Persistence & Privilege

always is false and model invocation is allowed (normal). The skill asks the operator to enable a sessionMemory config flag; this is a reasonable configuration change for its purpose but does alter agent defaults. The skill does write/append files to disk (MEMORY.md and daily logs) as its primary action, which is expected behavior.

Assessment

This skill appears coherent and light-weight, but be aware it will: (1) instruct enabling a sessionMemory agent config flag, which changes agent behavior platform-wide for agents that inherit defaults, and (2) write/append files (MEMORY.md and memory/YYYY-MM-DD.md) to disk as its primary action. Review where those memory files will be stored, ensure they are included/excluded from backups as you prefer, and periodically audit them for any sensitive content despite the skill's 'do not store secrets' guidance. If you need stricter privacy, do not enable the sessionMemory config or keep memory storage on an isolated path.

OpenClaw Memory Curator

A session-first memory system for OpenClaw.

It exists for one reason: important knowledge must survive session compaction without bloating the context window.

---

TL;DR (for humans)

• Session memory = temporary (RAM)
• Disk = source of truth
• Decisions & preferences → MEMORY.md
• Daily work → memory/YYYY-MM-DD.md
• This skill saves durable knowledge before compaction
• Retrieval always happens via memory_search → memory_get

If something matters later, write it to disk.

---

⚠️ CRITICAL REQUIREMENT

Session memory indexing must be enabled.

Enable Session Memory

CLI

clawdbot config set agents.defaults.memorySearch.experimental.sessionMemory true

JSON

{
  "agents": {
    "defaults": {
      "memorySearch": {
        "experimental": { "sessionMemory": true },
        "sources": ["memory", "sessions"]
      }
    }
  }
}

---

Mental Model (read this once)

OpenClaw memory has three layers. Confusion usually comes from mixing them up.

1. Session Memory (RAM)

• Lives in the current conversation
• Automatically compacted
• Indexed for retrieval
• Never reliable long-term

👉 Treat as short-term thinking space.

---

2. Daily Logs (memory/YYYY-MM-DD.md)

• Append-only
• What happened today
• Commands, edits, short-lived issues

👉 Treat as a work log, not a knowledge base.

---

3. Long-Term Memory (MEMORY.md)

• Curated
• Small
• High-signal only
• Indexed and retrievable

👉 Treat as facts the agent must not forget.

---

When to Write Memory (simple rules)

Write to MEMORY.md if it would still be true next week.

Examples:

• Decisions
• Preferences
• Invariants
• Policies

Write to daily logs if it helps understand today.

Examples:

• Refactors
• Experiments
• Temporary blockers

If unsure: write to daily log first, promote later.

---

Pre-Compaction Flush (why this exists)

Before OpenClaw compacts the session, it triggers a silent reminder.

This skill uses that moment as a Save Game checkpoint.

What happens:

1. Durable knowledge is extracted
2. Daily notes are written to today’s log
3. Durable items are promoted to MEMORY.md
4. Agent replies NO_REPLY (user never sees this)

This prevents knowledge loss without interrupting you.

---

Durable Memory Format (MEMORY.md)

Use IDs and tags so search works reliably.

## DEC-2026-02-04-01
type: decision
area: memory

Decision:
Session memory is retrieval-only. Disk is the source of truth.

Reason:
Session compaction is lossy. Disk memory is stable.

ID prefixes

• DEC – Decisions
• PREF – Preferences
• FACT – Durable facts
• POLICY – Rules / invariants

---

Retrieval Strategy (how agents should recall)

1. Use memory_search (max ~6 results)
2. Pick the best 1–2 hits
3. Use memory_get with line ranges
4. Inject the minimum text required

This keeps context small and precise.

---

Agent Playbook (rules for agents)

• Prefer disk over RAM
• Prefer MEMORY.md over daily logs for facts
• Use search before asking the user again
• Never copy raw chat into memory
• Write memory explicitly, do not assume it sticks

---

Anti-Patterns (do not do these)

• ❌ Copy chat transcripts into memory
• ❌ Store secrets or credentials
• ❌ Treat daily logs as long-term memory
• ❌ Overwrite memory files instead of appending
• ❌ Store speculation as fact

---

Privacy Rules

• Never store secrets (API keys, tokens, passwords)
• Ignore anything inside <private>...</private>
• If sensitive info exists: store only that it exists, not the value

---

Retention & Cleanup

Default: no deletion

• Disk is cheap
• Recall quality is expensive

Optional:

• Move old daily logs to memory/archive/YYYY-MM/
• Only prune after durable knowledge is verified

---

Usage (human-friendly)

Examples that work well:

• “Store this as a durable decision.”
• “This is a preference, remember it.”
• “Write this to today’s log.”

---

Design Philosophy

• Disk is truth
• RAM is convenience
• Retrieval beats retention
• Fewer tokens > more tokens
• Memory should earn its place