memory-distiller 🧠

The subconscious of your OpenClaw agent. Automatically distills fleeting conversation moments into permanent wisdom.

Most agents wake up blank. memory-distiller changes that — it automatically identifies what's worth remembering during each conversation, writes it to persistent memory files, and creates a true learning loop. Your agent gets smarter every session.

Why You Need It

OpenClaw natively has MEMORY.md and memory/YYYY-MM-DD.md, but updating them requires manual effort.

The problem: You correct the agent, it says "got it", and makes the same mistake next session.

memory-distiller closes this loop:

• - Automatically scans conversations for "worth remembering" signals
• Applies a quality gate to filter out one-time, temporary information
• Writes structured entries to memory files automatically

How It Fits With proactive-agent

One line: proactive-agent owns the present. memory-distiller owns the future.

Trigger Conditions

Auto-Triggers (scan every user message for these signals)

Manual Trigger

When the user says any of the following, immediately distill the session:

CODEBLOCK0

Quality Gate (All 4 Must Pass)

Before writing anything, check these 4 gates:

1. 1. Durability — Will this still be valuable in 24 hours? (Skip one-time context)
2. Generality — Is this a reusable rule, or a one-off special case? (Prefer rules)
3. Novelty — Does MEMORY.md already contain this? (Avoid duplicates; update if stale)
4. Actionability — Can this guide future behavior? (Skip vague impressions; only concrete rules)

Don't record (examples):

• - "Today we researched LangChain" → one-time, no guidance value
• "User is in Shanghai" → already in USER.md, duplicate

Do record (examples):

• - "clawhub install only accepts slugs, not GitHub URLs" → actionable rule
• "Feishu chat does NOT render Markdown — long reports must be sent as attachments or doc links" → prevents repeated mistakes

Memory Write Format

Where to Write

Entry Format

CODEBLOCK1

Example:

CODEBLOCK2

Execution Flow

CODEBLOCK3

Behavioral Rules

When to Notify the User

• - Auto-triggered → Write silently, do NOT say "I've noted that" — don't interrupt the flow
• Explicitly triggered → Confirm with one line: "✅ Noted." + brief summary of what was recorded

When NOT to Record

• - User says "suppose...", "hypothetically...", "for example..." → hypothetical, skip
• System errors, network timeouts → environment issues, not learnable rules
• Content already fully documented in MEMORY.md → don't duplicate; update if stale

Privacy Filter

Never write the following to any memory file, even if the user asks:

• - Passwords, tokens, API keys
• Personal identification info (ID numbers, bank accounts, etc.)
• Sensitive information about third parties

Installation

CODEBLOCK4

Zero configuration. Restart your OpenClaw session after installation.

Recommended Pairing

Together: never lose present details, never repeat past mistakes.

Design Philosophy

Inspired by Claudeception (a Claude Code self-learning plugin), redesigned from scratch for OpenClaw's architecture and memory model.

Core belief:

An agent should never be tripped up by the same problem twice.

Every correction is a learning opportunity. memory-distiller ensures none of them go to waste.

Academic foundation: Voyager (2023), CASCADE (2024), SEAgent (2025), Reflexion (2023) — all point to the same conclusion: agents that persist and reflect on their learning dramatically outperform those that start fresh every time.