OpenClaw's Founder Says "Try This Memory Plugin" — It Instantly Goes Viral

OpenClaw agents forget things. Everyone knows it. You spend 30 minutes setting up a complex workflow, the context window fills up, compaction kicks in — and your agent forgets half of what you discussed.

Peter Steinberger, OpenClaw's creator, just posted about a fix. His recommendation: Lossless Claw, a community plugin that replaces OpenClaw's built-in memory compaction with something that actually works. The post hit 277K+ views and 3,200+ likes in hours.

"If you are annoyed that your crustacean is forgetful after compaction, give Lossless Claw a try!" — Peter Steinberger

Here's what it does, why it matters, and how to set it up on MyClaw.ai in under a minute.

The Problem: OpenClaw's Memory Has a Hard Ceiling

Every AI model has a context window — a limit on how much text it can process at once. Claude Opus 4.6 tops out at 200K tokens. Sounds like a lot, until you realize:

📝 System prompts, memory files, and skill instructions eat 30-50K tokens before you even say hello

💬 A 50-message conversation easily hits 150K tokens

🔄 When the window fills up, OpenClaw's built-in compaction kicks in — and deletes older messages permanently

This is sliding-window compaction. It's simple, it's fast, and it's lossy. Once those messages are gone, they're gone. Your agent literally cannot remember what you discussed 20 minutes ago.

The result? You repeat yourself. Your agent makes the same mistakes twice. Workflows break because context vanished mid-execution.

What Lossless Claw Actually Does

Lossless Claw takes a fundamentally different approach. Instead of throwing away old messages, it:

🗄️ Persists every message in a local SQLite database — nothing is ever deleted

🌳 Builds a DAG (directed acyclic graph) of summaries — older messages get summarized, then summaries get summarized into higher-level nodes

🔍 Gives agents recall tools — lcm_grep to search history, lcm_expand to drill into any summary and recover original details

🛡️ Protects recent context — the last 32 messages are never compressed, ensuring smooth conversational flow

Think of it like this: regular compaction is a paper shredder. Lossless Claw is an archive with a search engine.

The Technical Details That Matter

Lossless Claw is based on the LCM paper and implements what the authors call "Lossless Context Management." Here's how the key parameters work:

🎯 Context threshold: 75% — Compaction triggers when context hits 75% of the model's window, not 95%. This leaves headroom for the model to actually think, instead of gasping for tokens at the last second

📊 Multi-layer summarization — Raw messages → leaf summaries → condensed summaries → top-level summaries. Each layer compresses further while maintaining links back to source material

🔗 DAG structure — Every summary knows which messages or sub-summaries it came from. This means an agent can ask "what did we discuss about the Ghost API?" and trace the answer back through the DAG to the exact conversation

⚡ Incremental compaction — Only new messages get processed each cycle, not the entire history. This keeps the per-turn overhead manageable

Real-World Impact

Here's what changes when you install Lossless Claw:

Before (stock OpenClaw):

• Agent forgets workflow details after long conversations
• You repeat instructions every few hours
• Complex multi-step tasks break when context compacts mid-execution
• No way to recover compacted information

After (Lossless Claw):

• Agent can recall any detail from any conversation using lcm_grep
• Multi-day projects maintain full context through the DAG
• Compaction happens earlier and smoother (75% vs near-100%)
• All raw messages persist in SQLite — nothing lost, ever

One user in Peter's thread asked whether this breaks context caching. Peter's response: "it's not default for a reason, it's a community plugin to explore new ideas!" The trade-off is real — each compaction cycle costs LLM tokens to generate summaries — but for agents running complex, stateful workflows, the benefit far outweighs the cost.

Install It on MyClaw.ai in 60 Seconds

If you're running OpenClaw on MyClaw.ai, installing Lossless Claw takes one command through your agent's chat:

openclaw plugins install @martian-engineering/lossless-claw

Your gateway restarts automatically, and you're running with DAG-based memory that never forgets. No SSH, no config files, no Docker — just tell your agent to install it.

The plugin works seamlessly alongside existing memory systems — your MEMORY.md files, daily logs, and skill configurations all continue working. Lossless Claw doesn't replace manual memory; it supplements it with automatic recall.

The combination is powerful:

🧠 MEMORY.md = what your agent should know at startup (decisions, preferences, contacts)

🗄️ Lossless Claw = what your agent can recall on demand (conversation details, exact instructions, debugging steps)

Together, your agent gets both fast startup context and deep recall — the best of both worlds.

Should You Install It?

If your OpenClaw agent runs simple, short conversations — probably not worth it. Stock compaction is fine for quick tasks.

But if you:

✅ Run multi-day projects where context matters

✅ Use your agent as a persistent assistant (not a one-shot tool)

✅ Are tired of repeating yourself after compaction

✅ Need your agent to remember what it built, deployed, or configured

Then Lossless Claw is exactly what you need. Peter Steinberger wouldn't recommend it to his 280K+ followers if it wasn't worth trying.

MyClaw.ai — the #1 OpenClaw host, the best way to run OpenClaw. Your agent's memory just got an upgrade.