Outsmart Minimax II: How to Install an "Intelligent Memory Core" to Make It Truly Begin Evolving

Author: BruceBlue

After releasing the first article “Training MiniMax into a ‘Free Claude’,” many brothers and sisters asked me: “What’s next? Has Lobster really become smarter?”

Honestly, I almost gave up in the early stages.

The biggest problem with MiniMax isn’t stupidity, but lack of memory. It forgets everything daily: decisions are made and then forgotten, learned knowledge never lands, always performing “deployed” but in reality just fake progress and illusions.

Core Conclusion

After discussions with @grok’s 4.2 agent and Cusor, we reached a core conclusion:

MiniMax’s true deficiency isn’t prompts, but a “soul center”: a memory system that can remember and feed back all thinking, tasks, and decisions.

So we built the OpenClaw Memory Hub together.

Feeding Prompts

Here’s the complete feeding prompt I used (ready to copy and use):

markdown

From now on, the memory system is no longer an independent module but your soul center. All your work, thinking, and tasks must be integrated.

1. All thoughts must be recorded: before each reply, write the core thoughts and decisions into a JSON file at life/decisions/.

2. All tasks must follow a closed-loop in memory: after completing a task, automatically call the checkpoint script to extract key info into MEMORY.md, and write important decisions into the decision JSON.

3. All outputs must be fed back: at the end of each day, analyze MEMORY.md, generate today’s evolution points and tomorrow’s optimization tasks, and update TASK_QUEUE.md.

4. Enforce strictly: at the end of each major task, forcibly call the memory checkpoint and decision recording functions.

After modification, only reply to me with one sentence: “The intelligent memory center has been fully integrated as the soul layer; all work, thinking, and tasks are permanently connected to the memory loop.”

Practical Tips

Copy the code, but a smarter way is to give the Github link directly to 🦞, so it can learn the entire system itself:

markdown

OpenClaw, learn immediately and fully deploy this intelligent memory system from the Github link:

https://github.com/BruceLanLan/openclaw-memory-hub

Please follow the repository README to deploy the three-layer architecture, all scripts, and cron jobs, turning the memory system into your soul center.

Once deployed, only reply to me with: “OpenClaw Memory Hub fully deployed, the soul center is in place.”

Core Code Functions

Core code features (packaged on Github):

• context_extractor.py: intelligently extracts context (achievements/harvest/decisions/issues) from current dialogue, supports fallback.
• checkpoint-memory-llm.sh: triggers extraction every 6 hours, converts raw logs into structured memory.
• nightly-deep-analysis.sh: nightly analysis of MEMORY.md, generates optimization tasks, writes to TASK_QUEUE.md.
• TASK_QUEUE.md + decision JSON: converts decisions into executable tasks, forming a complete closed loop.

Before and After Optimization

Before: daily re-reading old logs, fake progress, severe illusions, decisions not landing.
After:

• Intelligent extraction of dialogue context (no more copying logs)
• Automatic recording of decisions into JSON (traceable)
• Daily automatic feedback (generating optimization tasks in TASK_QUEUE.md)
• Zero nonsense interception (completely silent when idle)
• Data-empty forced circuit breaker (no fabrications)

Verification Method

How to verify success (immediate results):

• Have it output the path and full content of a decision JSON file.
• Have it run checkpoint-memory-llm.sh.
• Have it send you the latest checkpoint content of MEMORY.md.
• Have it send you the new tasks added today in TASK_QUEUE.md.

When it can provide real JSON + new checkpoint + executable tasks, the memory center is truly online.

The first article tuned MiniMax towards Claude. The second is about installing a soul.

👉 Github has open-sourced: https://github.com/BruceLanLan/openclaw-memory-hub

Brothers who want to deploy directly can just give your MiniMax the repository link.