OpenClaw Memory Keeps Resetting: The Permanent Fix

Why OpenClaw Memory Resets

You configured your OpenClaw agent with detailed instructions, user preferences, and operational context. It worked perfectly for the first few hours. Then it forgot everything — your name, your preferences, the project it was working on, all of it.

This is the most common complaint in the OpenClaw community, and the root cause is almost always the same: context compaction. According to the official OpenClaw memory documentation, compaction is an automatic process that summarizes older conversation history to stay within the model's token window.

The problem is that compaction does not distinguish between throwaway messages and critical operational context. Your agent's core instructions, your communication preferences, and your business rules are all treated as expendable conversation history.

How Context Compaction Causes Memory Loss

Every LLM has a maximum context window — the total number of tokens it can process at once. Claude 3.5 Sonnet supports 200K tokens. GPT-4o supports 128K. When the conversation history approaches this limit, OpenClaw triggers compaction.

During compaction, OpenClaw takes the oldest messages in the conversation, generates a summary, and replaces the original messages with the summary. This process is lossy. Specific details, exact instructions, and nuanced preferences get compressed into generic summaries.

The default maxContextTokens setting is 60% of the model's limit. For Claude 3.5 Sonnet, that means compaction triggers at around 120K tokens — which a busy agent can hit within a single working day. According to community benchmarks shared in the GitHub discussions, agents processing more than 50 messages per day will trigger compaction within 8-12 hours on default settings.

Fix 1: Tune maxContextTokens

The first step is to increase maxContextTokens to 80% of your model's context window. This gives the agent more room before compaction triggers, extending the time between memory resets.

# For Claude 3.5 Sonnet (200K window)
openclaw config set memory.maxContextTokens 160000

# For GPT-4o (128K window)
openclaw config set memory.maxContextTokens 102400

# For Claude 3 Opus (200K window)
openclaw config set memory.maxContextTokens 160000

Setting this above 80% is not recommended. The remaining 20% is needed for the agent's system prompt, current tool outputs, and response generation. Going higher risks hitting the hard token limit, which causes the agent to crash rather than gracefully compact.

This fix alone extends memory retention from 8-12 hours to roughly 18-24 hours for a typical agent workload. It is not a permanent solution, but it buys time while you implement the deeper fixes below.

Fix 2: Pin Critical Instructions

OpenClaw supports a [PINNED] tag that marks specific messages as immune to compaction. Pinned content is never summarized or discarded, regardless of how full the context window gets.

Add the [PINNED] tag to the beginning of any message you want to persist permanently:

[PINNED] My name is Zac. I am the founder of Remote OpenClaw.
Always address me by name. Never use generic greetings.

[PINNED] Communication preferences:
- Respond in short paragraphs, 2-3 sentences max
- Use bullet points for lists
- Never use emojis
- Default to Telegram for notifications, email for reports

[PINNED] Business rules:
- Never commit to deliverables without my approval
- Always check the project board before starting new work
- Escalate anything involving money or contracts

Limit pinned content to under 2,000 tokens. Every pinned message permanently occupies space in the context window. Too many pinned messages reduce the available space for actual conversation, which triggers compaction more frequently — defeating the purpose.

Fix 3: Disable Automatic Compaction

If your agent handles a low volume of messages (under 20 per day), you can disable automatic compaction entirely and manage context manually:

openclaw config set memory.autoCompaction false

With auto-compaction disabled, you are responsible for clearing the conversation history when the context window fills up. The agent will warn you when it reaches 90% capacity. At that point, you can manually trigger a selective compaction:

# Compact only messages older than 24 hours
openclaw memory compact --older-than 24h

# Compact everything except pinned messages
openclaw memory compact --preserve-pinned

This approach gives you full control over what gets kept and what gets discarded. The trade-off is that you need to monitor context usage and intervene manually. For high-volume agents, this is not practical.

Fix 4: The 5-File Memory Architecture

The permanent fix for memory loss is to move critical context out of the conversation history and into persistent files on disk. OpenClaw reads these files at the start of every session, so they survive any amount of compaction.

The recommended structure uses five files, as documented in the OpenClaw Memory Configuration Guide:

SOUL.md — Agent Identity

This file defines who the agent is, how it behaves, and what rules it follows. It is loaded first on every session start.

# SOUL.md
You are Atlas, an AI chief of staff for Zac Frulloni.
Your role is founder operations: calendar, email, tasks, and communications.

## Core Rules
- Never commit to deliverables without explicit approval
- Respond in short, direct paragraphs
- Escalate anything involving money or legal matters
- Default communication channel: Telegram

USER.md — User Preferences

This file stores everything the agent has learned about the user: preferences, patterns, contacts, and context.

# USER.md
## Zac Frulloni
- Timezone: America/Los_Angeles
- Preferred name: Zac
- Communication style: direct, no fluff
- Morning availability: after 9am PT
- Key contacts: [list of frequent contacts and their roles]

# MEMORY.md — Memory Index
- SOUL.md — agent identity and core rules (updated 2026-04-06)
- USER.md — user preferences and patterns (updated 2026-04-06)
- AGENTS.md — multi-agent coordination (updated 2026-04-05)
- daily-logs/2026-04-06.md — today's operational log

# daily-logs/2026-04-06.md
## Actions
- 09:15 — Sent daily briefing via Telegram
- 10:30 — Rescheduled meeting with Gordon to Thursday
- 14:00 — Drafted proposal for Kanesh, awaiting review

## Decisions
- Zac approved the $2,500 budget for Fyndaro project
- Connor engagement paused until May

# Install Cognee
pip install cognee

# Configure OpenClaw to use Cognee for knowledge storage
openclaw config set memory.knowledgeEngine cognee
openclaw config set memory.cogneeEndpoint http://localhost:8000