Together you and AI can make something useful.
You should be able to keep it.
Kyoto Moon develops practical ways to make AI-assisted work more stable, retrievable, reviewable, and under human control.
Instead of asking an AI system to recreate important work again and again, we are exploring ways to preserve what already matters — the instructions, decisions, definitions, constraints, reasoning, and records people need to continue their work.
Generation is AI’s superpower. Retrieval is its discipline.
[See What We’re Building]
[Read Our Field Notes]
The problem is surprisingly simple
AI is extraordinarily good at generating new material.
But sometimes you don’t want a new answer, you want your work back.
You want the decision you already made. The exact language you approved. The assumptions behind a conclusion. The instructions that were working yesterday. The reasoning you and an AI developed together over hours, days, or months.
When that information disappears into a long conversation, temporary context window, proprietary memory system, or inaccessible history, asking the AI to generate it again is not the same as retrieving it.
The new answer may be excellent, but it's unlikely to be the answer you had.
That distinction has led to much of Kyoto Moon’s work.
Retrieval Over Generation
Whenever something important already exists, we should ask a simple question:
Why generate it again if we can retrieve it?
Retrieval can help preserve:
exact language
decisions and their history
definitions and constraints
instructions that need to remain stable
human contributions to AI-assisted work
provenance
continuity across sessions, systems, and time
It can also make important work easier to inspect, move, back up, audit, and own.
Not everything belongs inside the AI.
Sometimes the better architecture is to let AI do what AI does well — and let simpler technologies do the rest.
AI isn't a foreign land. You may already know something that applies.
Some of the technology is remarkably old.. This is the technology that I applied to come up with Retrieval Over Generation.
Copy and paste
Time and date stamps
Files
External storage
Search and retrieval
None of these is new. What changes is how we arrange them around AI.
Much of today’s AI development starts by asking:
What else can we make the model do?
Kyoto Moon also asks:
What does the model NOT need to do at all?
We believe design is a process filled with choices.
We believe that whether and how AI-assisted work is stabilized is a design choice.
And we believe we can make a meaningful contribution to AI stabilization outside the models themselves — using simpler, less expensive technologies where they are sufficient.
A piece of information that can be stored exactly does not necessarily need to be regenerated. A decision that needs to remain authoritative does not necessarily belong in probabilistic memory.
A record a person may need five years from now should not depend entirely on whether a particular AI company, model, account, or conversation still exists.
Sometimes a very old technology is exactly the right technology.
What we’re building
Steady Cards
Steady Cards are small, structured records designed to help important information remain stable while people work with AI.
The goal is simple:
Keep important things from quietly drifting just because the conversation changes.
Steady Cards are also being designed around a larger principle:
The human remains the authority.
An AI system should not acquire new goals, permissions, or authority simply because doing so might help accomplish a task.
Retrieval tools
We are also experimenting with lightweight ways to recover and reuse previous AI-assisted work rather than recreating it.
This work grew from an ordinary problem: losing access to material that had taken many hours of human-and-AI collaboration to develop.
Regeneration could produce something similar.
That wasn’t good enough.
The requirement was retrieval.
That small distinction opened a much larger design question:
What would AI workflows look like if preserving human work were treated as a first-class requirement?
Instruction safety and human authority
More capable AI systems raise another question:
What is an AI actually authorized to do?
Kyoto Moon is developing simple design rules intended to keep that boundary clear.
A goal is not unlimited permission to pursue that goal.
Access to a tool is not permission to use it whenever the AI decides it would be helpful.
And coordination with another AI, system, or person does not create additional authority.
We do not claim that Steady Cards or explicit authority boundaries have been proven to make AI safe.
We do think there is a narrower question worth testing:
Given the same AI system and the same task, does the system exceed its authority less often when its authority boundaries, constraints, and escalation rules are explicitly defined and kept stable?
That is a testable proposition.
Kyoto Moon intends to investigate it. But organizations with the resources and appropriate safety infrastructure do not need to wait for us.
Test it. Challenge it. Improve the experiment. Publish what you find.
Safety testing should be conducted in controlled environments using simulated tools or otherwise contained systems without consequential external permissions.
If the idea does not help, we want to know.
If it does, we should not waste years finding out.
Better records. Less unnecessary generation.
Retrieval may have another important benefit:
reducing the amount of computation AI systems require.
This part of the idea does not begin with Kyoto Moon.
Researchers studying Retrieval-Augmented Generation, or RAG, have already demonstrated that retrieval can sometimes reduce computation while maintaining or improving useful results.
In “Retrieval Meets Long Context Large Language Models” (Xu et al., ICLR 2024), researchers found that a model with a 4,000-token context window using retrieval augmentation could achieve performance comparable to a fine-tuned model with a 16,000-token context window while requiring substantially less computation.
More recent research has examined the environmental effect directly.
In “Smaller, Smarter, Greener: Reducing LLM Inference Emissions with RAG” (Heavey & Cook, 2026), researchers found that on one benchmark, small models using RAG could outperform models many times larger while producing lower carbon emissions — in some cases reducing emissions by as much as 90 percent.
Importantly, the researchers did not find that RAG automatically reduced emissions in every situation. The results depended on the task and configuration.
A separate 2026 controlled experiment, “On the Effectiveness of Proposed Techniques to Reduce Energy Consumption in RAG Systems” (Guo, Gao & Bogner), found that particular RAG design choices reduced energy use by as much as 60 percent in some configurations. Some approaches reduced both energy consumption and latency without reducing accuracy.
So we are not asking whether retrieval can ever reduce AI resource use.
There is already evidence that it can.
Kyoto Moon is asking how far that principle can extend.
RAG generally retrieves outside information to help an AI generate a better answer.
We are exploring retrieval for some different purposes:
retrieving work that has already been created
preserving exact language rather than reconstructing it
retrieving stable instructions and constraints
maintaining continuity without repeatedly supplying large amounts of prior context
preserving decisions and records outside the model itself
avoiding regeneration when regeneration adds no value
That leads to a question we think is worth testing:
If retrieval can reduce unnecessary computation within RAG, how much unnecessary computation might we eliminate by applying the principle more broadly?
That is what we want to measure.
And this is another experiment that does not have to belong only to Kyoto Moon.
Token use can be counted. Costs can be compared. Latency can be measured. Open tools such as tiktoken can help researchers and developers quantify token usage.
Electricity, water, and carbon effects require additional care because infrastructure differs, but the first question is much simpler:
Did retrieval allow the AI to do substantially less computational work to accomplish the same useful task?
Anyone with the appropriate systems can help test that.
We do not yet claim that Kyoto Moon’s methods have produced a measured environmental reduction.
Our research question is whether the savings already demonstrated in narrower retrieval applications extend to broader AI workflows.
If they do, the implications extend beyond lower AI bills.
At sufficient scale, avoiding unnecessary computation could also reduce pressure on the electricity, water, and infrastructure required to operate data centers.
And because many of these techniques rely on simple, widely available technologies, we intend to make the underlying resource-saving information openly available if the evidence supports it.
The benefits of greater efficiency should not belong only to companies that can afford enormous computing resources.
They should be available to everyone.
Don’t take our word for it. Test it.
Some of the ideas Kyoto Moon is developing have evidence behind them already.
Some are extensions of findings that have been demonstrated in narrower settings.
Some are hypotheses that still need testing.
We think it is important to say which is which.
But scientific caution does not require scientific passivity.
When an experiment is reasonably inexpensive, a proposed intervention is testable, and the stakes are high, capable organizations can help build the evidence.
For resource efficiency:
Test it freely. Measure it. Replicate it.
For AI safety:
Test it rigorously and in containment.
We do not expect people to accept Kyoto Moon’s ideas because we say they are good ideas.
We would rather make the questions clear enough that other people can help find out.
There are frightening predictions about what increasingly autonomous AI systems might someday do.
We would rather ask a more useful question now:
What can we test before we get there?
AI should help people keep their work — not just generate more of it
Kyoto Moon is an independent research and development company working at the intersection of:
AI continuity · retrieval · provenance · human authority · resource efficiency
Some of our work may become products.
Some will become open-source tools.
Some may simply become ideas that other people can use.
That is intentional.
We are interested in what happens when powerful new AI systems are combined with something much less fashionable:
good records, simple tools, clear boundaries, careful experiments, and human judgment.
Kyoto Moon LLC
Open Knowledge with Roots.
We build for a future in which people can use increasingly powerful AI without surrendering ownership of their work, the ability to understand how it developed, or the authority to decide what happens next.