About Openclaw memory, you just need to see this。
NO PAPERS, NO WORDS. AFTER READING THIS, YOU'LL KNOW MORE ABOUT AI THAN 901 TP3T。
I'll take itOpenClaw TAKE A LOOK AT HOW THE ENGINEERS ARE GIVING AI ONE STEP AT A TIME
From the roughest of options to the most forward line of thinking。
AI, WHY AMNESIA
LET'S FIND OUT THE BASIC QUESTION: WHAT IS AI'S BRAIN。
When you talk to ChatGPT, Claude, these AIs, you probably think it's like a human being, and you remember what you talk about。
Not really。
AI'S BRAIN IS NOT A HARD DRIVE, IT'S A DESK. THE SIZE OF THE DESK IS FIXED, AND EVERY WORD YOU SAY TO IT, AND EVERY WORD IT RETURNS, IS A PIECE OF PAPER ON THE TABLE. WHEN THE TABLE IS FULL, THE NEW PAPER WILL NOT BE RELEASED。
This desk, technically called the context window。
Now the big model, the table's not small, Claude's table probably opens up a whole Harry Potter. That sounds big. But the problem is that every word you talk about itself is on the table. The longer the conversation takes, the thicker the record of the dialogue on the table, the less room is left for real work。
EVEN WORSE, AI DOESN'T HAVE A DRAWER. HUMANS CAN PUT WHAT IS NOT NEEDED FOR THE TIME BEING IN THE DRAWER AND TURN IT OUT WHEN THEY NEED IT. BUT FOR AI, THERE IS NOTHING ON THE TABLE。
When the table is full, old things are pushed down and left empty。
IT'S NOT AI STUPID, IT'S STRUCTURAL CONSTRAINTS. JUST LIKE YOUR BROWSER, IT'S THIS BIG, ROLL DOWN AND IT'S INVISIBLE. THE DIFFERENCE IS, YOU CAN STILL TURN BACK, AI CAN'T. GET THE HELL OUT OF HERE, IT DOESN'T EXIST FOR IT。
What if the table's full? The most intuitive idea is to rub old paper, press it, and make room。
OpenClaw's first plan, that's all。
Option 1: Scroll old notes into paper
OpenClaw is an open-source AI programming assistant platform that you can interpret as an AI programmer management system。
It initially dealt with memory in a simple and brutal manner:
WHEN THE DIALOGUE IS FULL OF DESKTOP 92%, THE SYSTEM AUTOMATICALLY TRIGGERS COMPRESSION. KEEPS THE MOST RECENT THREE ROUNDS OF DIALOGUE, WITH ALL THE PREVIOUS CONTENTS PACKAGED FOR ANOTHER AI SUMMARY. FINISH THE SUMMARY AND POST IT BACK TO THE DESKTOP。
What about the conversation? Delete permanently. Can't find it back。
Just like your desk's almost full, you put 20 pages in front of you and asked your colleagues to write you a summary and throw that 20 pages into the shredder. The table is clear, but next time you want to remember how much that parameter was, I'm sorry, it's not back in the shredder。
The programme has three questions:
FIRST, THE SUMMARY WILL LOSE DETAILS. YOU TOLD AI THAT THE VARIABLE NAME WAS CHANGED, 20 ROUNDS LATER THE INFORMATION WAS COMPRESSED, AND AI USED THE OLD NAME BACK。
SECOND, IRREVERSIBLE. THE ORIGINAL DIALOGUE WAS PERMANENTLY LOST AND IT WAS IMPOSSIBLE TO FIND OUT WHY OPTION A HAD BEEN CHOSEN。
Thirdly, the pressure gets worse. If the compression triggers more than one time, it is a summary of the summary, and the information appears to be reproduced over and over again and is becoming less visible。
It works, but it's thrown every time it's compressed. The longer we talk, the more we lose。
Is there any way to compress, but not to throw things away
Programme II: Establishment of a layered archive
A strong team in the OpenClaw community made a second-generation program called Losles Cloud. The core commitment is intact, and the original message is always kept and none is deleted。
It's very clever. Do not delete the original and develop a thumbnail system。
Imagine Google Earth. You won't load every inch of the world's street view at the same time. You see: Earth's panorama, zoom in on the country, zoom in on the city, pull into the street。
This is what Losles Cloud did:
- Bottom (street view): Every original conversation, one word at a time, always in a local database
- Mid-level (city map): Summaries of relevant dialogues
- Top level (country map): global summary of the entire dialogue
Only the top-level summary is displayed on the desktop. Need details? Just turn it down. It's always there。
Much better than option one. At least don't throw things away。
But deep down, three questions came up:
FIRST, A NEW DIALOGUE WINDOW IS A MEMORY LOSS. THE ARCHIVE SERVES ONLY ONE DIALOGUE. YOU TURN OFF THE WINDOW AND OPEN A NEW, AI, ANOTHER WHITE PAPER. THE PITS THAT HAD BEEN STEPPED ON BY THE PREVIOUS PROJECT HAD YET TO BE STEPPED AGAIN。
Secondly, the desktop will be full. Although the originals were entered into the warehouse, the digest itself took up space. The longer the dialogue, the thicker the summary. Like more and more of your books, it's half a table。
THIRD, THE WAREHOUSE WAS TOO SLOW. TO FIND A SPECIFIC ORIGINAL RECORD, IT HAS TO BE CARRIED OUT FROM THE TOP LEVEL, THE PROCESS HAS TO CALL AI REPEATEDLY, AND YOU MAY HAVE TO WAIT TWO MINUTES TO GET THE RESULTS。
It did. But when you need it to remember, it turns too slow. And each time a new conversation takes place, it amounts to a new empty house。
Here comes a more fundamental question:
Both options are thinking of the same thing, how to put more and more into a limited table. One is to throw away, one is to compress。
But is it possible that the thought itself is wrong
Option three: Don't pile it on the table. Just take it as needed
That's MemOS' idea. I've been working on my own, and I've seen it Agent Remember the most complete solution to this. It's a plugin for OpenClaw, open source free, 100% locally run, all data on your own computer, and zero upload。
MemOS can't compress no more, and it asks a completely different question: Why does everything have to be on the table
What's the first two? Like you're on a business trip, stuffing all your clothes in a suitcase. One is to throw out a few pieces (Legacy) and the other is to compress with a vacuum (Lossless Cloud). The suitcase is still so big that it will be filled sooner or later。
MemOS said, "Don't carry the suitcase." Put your clothes in the closet and go out with what you're wearing today。
How did you do that
Full records, but not on the table. Every sentence you talk to AI, every operation, is automatically stored in an independent local database. Don't pick or pick, save it all. However, these records do not occupy the desktop, and they remain silent in the database. If you've used OpenClaw's own memory function, MemOS supports a one-key migration, and the context that you have accumulated is not in vain。
Retrieval on demand, only relevant. AI will not move all history to the desktop when it receives a new assignment, but, like the search engine, will search the database according to the current assignment, and will transfer only a few of the relevant entries. Last week, Nginx? This mission has something to do with Nginx, and that memory has appeared. Does not matter an inch of the desktop。
What's the impact of this? The MemOS team ran two teams。
The first was the test of the open data set LOCOMO: token consumption fell directly by over 72%, while the accuracy rate increased by 33.5%. Saved money and answered more accurately because the context was filled with information and less noise。
The second group is a real engineering scene, with complex development tasks carried out through multiple sessions: dialogue rounds dropped from an average of 116 to 54, total token consumption dropped from 2.2 million to 1.12 million, and 49% was cut off. The speed of mission completion is 2.15 times higher。
Previously, desktop consumption had been proportional to the amount of talk, and was now linked to what was needed for the current mission and was largely constant。
Automatic summary, enrichment experience. For each mission performed, the system automatically generated a structured summary: what was done, what was concluded, what was stepped on. Two hours of debugging, condensed into a card. Next time you get a similar assignment, just pull out the card。
Automatic evolution of skills. This is the worst. The system will extract reusable Skills from a recurring pattern. For example, you let AI handle the CSV file in the same way three times in a row, and it automatically draws up a rule, and next time it goes straight. And these Skills aren't static, and it's a better way to upgrade itself. It's like when a new employee goes from asking everything to having his/her own working method, and it's going on over time。
Cross-dialogue, cross-Agent sharing. The first two options, a new dialogue window for AI, are memory loss. MemOS won't. At the beginning of the new dialogue, the system automatically retrieves relevant memories, and the experience of the previous project is of direct use. If you have multiple Agents, like one writing code, one running test, they share the same memory library, and the whole team can see it. MemOS calls this MemOS Hub, which is essentially a team knowledge hub。
We'll look at the three options together
The three programmes correspond to three distinct perspectives:
The programme felt that memory was a burden, too much to compress。
Option 2 feels that memory is an asset, cannot be lost and must be preserved。
Option 3 feels that memory is a capacity, not only for storage, but also for use and evolution。
Want to try Memos
MemOS is an OpenClaw plugin that is free of charge and can be installed with a line of commands。
Mac / Linux:
curl-fsSL https://cdn.memtensor.com.cn/memos-local-openlaw/install.sh bashWindows (PowerShell):
powershell-c 'irm https://cdn.memtensor.com.cn/memos-local-openclaw/install.ps1 iex'Plugin home page and document:
https://memos-claw.openmem.net/
A large-scale model can be used without additional databases and external dependence。
REMEMBERING DETERMINES HOW FAR AI CAN GO
Back to the beginning of the commitment。
AN AI WITHOUT MEMORY, SMART ENOUGH TO BE AN INTERN ON HIS FIRST DAY OF SERVICE FOREVER. IT'S BEEN A GREAT DAY, BUT YESTERDAY'S TEACHINGS WERE FORGOTTEN。
It was given a vague summary note. Option 2 gave it a hard-to-turn, but slow-to-turn archive. Option 3 gives it a truly functional knowledge management system that is capable of being, being, being, being, learning and evolving。
Agent's memory is early, but the direction is clear. The future AI will not only be smarter, but will also remember。
AN AI THAT WRITES, IS AN AI THAT YOU'RE WILLING TO WORK WITH FOR A LONG TIME。
