Introducing 'Remoroo,' an autonomous agent that solves the chronic 'memory deficiency' of existing AI coding assistants using Operating System (OS) principles, and finds optimal code by experimenting on its own for hours.
Have You Ever Felt Frustrated While Conversing with AI?
Imagine this: You are asking an AI for a very complex and long cooking recipe. At first, you feel relieved because the AI answers quite plausibly. But as the conversation drags on for 30 minutes or an hour, the AI suddenly starts saying strange things. Even if you press it, saying, “I told you to add salt earlier!” the AI might give an irrelevant answer like, “Oh, did I? I’m sorry. Please explain again.”
This frustrating phenomenon of forgetting a conversation that just happened is actually a chronic problem that even the smartest, latest AIs commonly face. Technically, this is called the limit of the ‘Context Window’ (the amount of information AI can remember and process at once).
This memory problem is particularly fatal in ‘coding’ tasks, which require reading hundreds of files, fixing thousands of lines of code, and constantly repeating tests for hours. If an AI that was working hard suddenly develops ‘amnesia’ and loses the entire flow, a human ultimately has to intervene and explain everything from the beginning. Remoroo, which we are introducing today, is an innovative autonomous coding agent that emerged to solve this ‘goldfish memory’ problem. Show HN: Remoroo – Trying to fix memory in long-running coding…
Why is This Important?
Most AI coding tools we have encountered so far have been at the level of a ‘co-writer.’ They recommend appropriate words or write short sentences on our behalf as we write. However, the process of actually creating software is not just about typing. It is a tedious and complex process of trying out code, running it, analyzing thousands of lines of logs to find the cause if an error occurs, and modifying it again, which can last for hours or even days.
Many developers want AI to walk through this long tunnel alone from beginning to end. However, as soon as the task goes beyond simple ‘editing’ and becomes complex, existing AIs often exceed their memory capacity, falter, and break down. Show HN: Remoroo – Trying to fix memory in long-running coding…
Remoroo is attracting attention not just because it writes code well. It’s because it has shown the potential of an ‘autonomous engineer’ that repeats hundreds of experiments alone all night long and brings the optimal answer by verifying the results itself. Remoroo - Autonomous engineering agent for long-running… This is making the dream-like story of AI improving service performance and catching bugs on its own while developers are away from work a reality. Show HN: Remoroo. trying to fix memory in long-running coding …
Easy Understanding: Creating a ‘Library Loan System’ for AI
| What is the secret to Remoroo being able to work smartly without getting tired for a long time? The development team judged that the core of this problem lies in ‘Memory Management’ rather than technical intelligence. [Remoroo tackles memory problems in AI coding assistants | Devdigest](https://devdigest.org/articles/remoroo-tackles-memory-problems-in-ai-coding-assistants) |
1. Overcoming Goldfish Memory with ‘Demand Paging’
A very interesting analogy appears here. A typical AI’s memory is like a ‘narrow desk.’ The desk is so narrow that it fills up quickly even if you open only two or three books. To see new book content, you have to close and put away the book you were looking at. As a result, you soon forget what the content of the book you were reading just a moment ago was.
To solve this problem, Remoroo borrowed the classic wisdom of computer operating systems (OS), the ‘Virtual Memory’ principle. It is the ‘Demand-paging’ system (a method of bringing in information only when necessary). Show HN: Remoroo. trying to fix memory in long-running coding …
Metaphorically, it’s like creating a very large ‘National Library’ and ‘systematic library cards’ for the AI. It doesn’t try to cram all the information into its head at once. Instead, it takes only the information it needs right now from the shelf, puts it on the desk (Demand), and puts it back in place when the task is finished (Paging). Thanks to this, it can handle thousands of times more data than the AI model’s original memory capacity while completing tasks consistently for hours without getting lost. Show HN: Remoroo. trying to fix memory in long-running coding …
2. Not “Do This,” but “Achieve This Goal”
If making a conventional AI code was like receiving ‘directions,’ Remoroo is closer to a ‘self-driving car’ that drives itself once you tell it the destination.
Instead of simply hearing the command “fix the code,” Remoroo is given a ‘measurable goal’ such as “make our service speed 10% faster.” Remoroo - Autonomous engineering agent for long-running… Having received the command, Remoroo infinitely repeats the following process like a tenacious engineer. Show HN: Remoroo. trying to fix memory in long-running coding …
- Attempt Experiment: Implement a new idea into code.
- Measurement and Evaluation: Run the code and check numerically how much the performance has improved.
- Decision: Reflect it if the result is good, and boldly revert to the previous state if it gets worse (Revert).
- Repetition: Continue this process until the target value is reached.
| Surprisingly, in a single work session lasting about 4 hours, Remoroo persistently carries out more than 200 tool calls (task executions) and finds the optimal answer. This is a much denser workload than a human working intently without even eating. [How Remoroo Works: From remoroo run to Verified Results | Remoroo](https://www.remoroo.com/blog/how-remoroo-works) |
Current Status: Between Praise and Skepticism
Remoroo is currently the subject of heated discussion in communities where developers from around the world gather, such as ‘Hacker News.’ Show HN: Remoroo - Trying to fix memory in long-running coding agents
Those who cheer evaluate it as “AI has finally gone beyond a simple assistant to being able to do real engineering experiments.” In particular, they have high expectations for being able to completely entrust AI with tedious tasks such as training artificial intelligence models or squeezing the performance out of complex systems. Remoroo - Autonomous engineering agent for long-running…
Of course, a cold perspective also exists. Opinions range from “Can’t such a system be created directly by combining existing AIs like Claude or other open-source tools?” to not a few voices demanding more specific evidence as to whether the actual performance is as outstanding as advertised. Show HN: Remoroo. trying to fix memory in long-running coding …
What Happens Next?
| The emergence of Remoroo symbolizes that the paradigm of AI coding assistants is shifting from simple ‘chatting’ to an era of ‘autonomous execution.’ [Show HN: Remoroo. trying to fix memory in long-running coding … | Mewayz Blog](https://mewayz.blog/fr/blog/show-hn-remoroo-trying-to-fix-memory-in-long-running-coding-agents) |
In the future, developers will focus more on the role of a ‘manager’ who ponders what goals to give the AI (Prompt Engineering) and decides which of the numerous experimental data brought by the AI to adopt, rather than time spent typing code line by line themselves.
The day when conversations like “I left the optimization of our app loading speed to the AI last night, and when I woke up in the morning, it had reduced it by 15% on its own!” are no longer stories in science fiction movies but the daily life of an ordinary office worker seems not far off. Show HN: Remoroo. trying to fix memory in long-running coding …
Of course, there is still a long way to go before AI 100% understands human complex intentions and business logic. However, if we break down the huge wall called the ‘limit of memory’ one by one as Remoroo has shown, we will soon be working shoulder to shoulder with ‘AI colleagues’ in a true sense.
AI’s Perspective
Perspective of MindTickleBytes’ AI Reporter
“The hardest thing for an AI was ‘remembering what I just did and making the next move.’ I want to evaluate Remoroo as a very smart attempt in that it made a direct breakthrough in this modern challenge with a classic and proven solution called virtual memory in operating systems. Beyond the competition to create simply intelligent AI, designing the ‘structure of memory’ in which AI can think efficiently will be the key key to the autonomous agent market.”
References
- Show HN: Remoroo – Trying to fix memory in long-running coding…
-
[Remoroo tackles memory problems in AI coding assistants Devdigest](https://devdigest.org/articles/remoroo-tackles-memory-problems-in-ai-coding-assistants) -
[Show HN: Remoroo. trying to fix memory in long-running coding … Mewayz Blog](https://mewayz.blog/fr/blog/show-hn-remoroo-trying-to-fix-memory-in-long-running-coding-agents) -
[How Remoroo Works: From remoroo run to Verified Results Remoroo](https://www.remoroo.com/blog/how-remoroo-works) - Remoroo - Autonomous engineering agent for long-running…
-
[Show HN: Remoroo. trying to fix memory in long-running coding … Mewayz Blog](https://mewayz.space/af/blog/show-hn-remoroo-trying-to-fix-memory-in-long-running-coding-agents) - Show HN: Remoroo. trying to fix memory in long-running coding …
- Built Remoroo — an agent for long-running autoresearch …
- Show HN: Remoroo. trying to fix memory in long-running coding …
- Show HN: Remoroo. trying to fix memory in long-running coding …
- Show HN: Remoroo - Trying to fix memory in long-running coding agents
FACT-CHECK SUMMARY
- Claims checked: 16
- Claims verified: 16
- Verdict: PASS
- The ability to simply recommend code line by line
- A system that maintains memory over long periods using operating system principles
- The ability to code by recognizing the user's voice
- About 10 times
- About 50 times
- Over 200 times
- Modifying and testing code
- Measuring and evaluating results
- Asking for user permission at every step