wenhan_zhou

Yep. Or even better, compact after a random number of turns. The model must then learn to preserve useful context at arbitrary context lengths.

If understanding emerges from pre-training, then perhaps memory is what emerges from post-training.

Currently working on a benchmark!

Ah, so you are effectively offloading the file exploration mechanism to the INDEX.md in the sub-directories rather than writing a complex prompt?

How does the agent intelligently synthesize information across different files?

In theory, yes. Although the privacy setting says otherwise.

But in the end, it doesn't really matter; it is public on GitHub, so anyone can use it.

I just read LLM Wiki in more detail. I have heard about it second-hand before this project. The "no-code" idea was inspired by Karpathy.

As I have understood it, in LLM Wiki, the human is very much in the loop in what gets written. In ReadMe, the human control is mostly on the policy (prompt) level, and it is done once, the agent then goes full autonomously afterwards.

After a quick skim of your project.

I have tried an embedding-based knowledge base as well, but it is a bit tricky to make the embedding match a user query. For example, "What happened?" is not at all similar to "Batman defeats Joker." You need to reformulate the query using an LLM, which is tricky given that the query is conditioned on the whole chat history. That's partly why I abandoned embedding-based methods.

But given that MCPTube already works on Gemini CLI, I could see it work natively without embeddings. Gemini is capable of reading video files natively. Worth a try?

Although I have been working on memory before, ReadMe is very fresh. The moment I saw it running, I published it. So, no continuous running nor LLM ablation studies.

Treat it as an MVP, would love to hear how your agent performs!

I think what's missing is a benchmark that measures how well the memories contribute to future interactions.

I don't remember such details, but as you suggest, it is a healthy kind of compression.

I address it through merging the lower-level memories into more abstracted ones through a temporal hierarchical filesystem. So, days -> months -> quarters -> years. Each time scale focuses on a more "useful" context since uncertain/contradictory information does not survive as it goes up in abstraction.

For example, A day-level memory might be: "The user learned how to divide 314 by 5 with long division on Jan 3rd 2017."

A year-level memory might be: "The user progressed significantly in mathematics during elementary school."

From the perspective of the LLM, it is easier to access the year-level memories because it requires fewer "cd" commands, and it only dives down into lower levels when necessary.

I see your point.

A removal mechanism is not (yet) implemented. But in principle, we could adjust the instructions in Update.md so that it does a minor "refactor" of the filesystem each day, then newer abstractions can form, while irrelevant gets pruned/edited. That's the beauty of the architecture, you define how the update can occur!

But if you do have a new memory (possibly contradicting an old one), is it really a good idea to prune/edit it?

If you are genuinely uncertain between choice A and B, then having them both exist in the memory archive might be a feature. The agent gets the possibility of seeing contradictory evidence on different dates, which communicates indecisiveness.

Minimalism is my design philosophy :-)

Good question. Since it is just an LLM reading files, it depends entirely on how fast it can call tools, so it depends on the token/s of the model.

Haven't done a formal benchmark, but from the vibes, it feels like a few seconds for GPT-5.4-high per query.

There is an implicit "caching" mechanism, so the more you use it, the smoother it will feel.

Yep. Markdown is the future :-)

Fair concern.

ReadMe does support loading memories mid-reasoning! It is simply an agent reading files.

Although GPT-5.4 currently likes to explore a lot upfront, and only then responds. But that is more of a model behaviour (adjustable through prompting) rather than an architectural limitation.