Outsmarting the Minimax II: How to equip it with an "intelligent memory center" and truly begin its evolution

Author: BruceBlue

After the first article in this battle of wits, "Turn the MiniMax into a 'Free Claude'," was published, many brothers and sisters asked me, "What happened next? Did the lobster really get smarter?"

To be honest, I almost gave up in the early stages.

MiniMax's biggest problem isn't its stupidity, but its lack of memory. It acts like it has amnesia every day: it forgets decisions as soon as they're made, it never puts what it learns into practice, and it's always putting on a show of "deployment" when it's all just fake progress and illusions.

Key conclusions

After discussing with @grok 's 4.2 agent and Cusor, I reached the following core conclusions:

What MiniMax truly lacks is not prompts, but a "central hub": a memory system that can remember and feed back all thoughts, tasks, and decisions.

So together we created OpenClaw Memory Hub

Feed Prompt

The complete feeding prompt I used at the time (can be copied directly):

Practical Tips

Copying the code is actually a smarter approach; just give the GitHub link to the program and let it learn the whole system on its own.

Core code functionality

Core code functionality introduction (packaged on Github):

• context_extractor.py: Intelligently extracts (achievements/gains/decisions/problems) using the current dialogue context, and supports a fallback mechanism.

• checkpoint-memory-llm.sh : Automatically triggers extraction every 6 hours, converting the raw logs into structured memory.

• nightly-deep-analysis.sh : Analyzes MEMORY.md every night, generates optimization tasks, and writes them to TASK_QUEUE.md.

• TASK_QUEUE.md + Decision JSON: Transforms decisions into executable tasks, forming a complete closed loop.

Comparison before and after optimization

Before optimization: Daily rereading of old logs, false progress, severe illusions, and decisions not being implemented. After optimization:

• Intelligent extraction of dialogue context (no longer copying logs)

• Decisions are automatically recorded as JSON (traceable).

• Daily automatic feedback (TASK_QUEUE.md generates optimized tasks)

• Zero-message interception (completely stops sending messages when silent).

• The circuit breaker will be forcibly triggered if the data is empty (this is no longer fabricated).

Verification method

How to verify success (immediate results):

• Make it output the path and full content of a decision JSON file.

• Run checkpoint-memory-llm.sh

• It will send you the latest checkpoint content of MEMORY.md.

• Let it send you the newly added tasks in TASK_QUEUE.md today.

When it can provide real JSON, a new checkpoint, and an executable task, it means that the memory center is truly online.

The first article is about sculpting the MiniMax in the direction of Claude. The second article is about giving it a soul.

👉Open source on GitHub: https://github.com/BruceLanLan/openclaw-memory-hub

If you want to deploy directly, just send the repository link to your MiniMax.