Show HN: Screen memory without screenshots, just text to Markdown

DDramatize about 6 hours ago 16 commentsRead Article on github.com

ZH version is available. Content is displayed in original English for accuracy.

It's a macOS menu bar app that reads the text of your focused window every few seconds through the Accessibility API. No screenshots, no video, or OCR. It writes plain markdown, one file per day, into a folder you choose.

Point Claude Code (or anything with file access) at the folder and ask "what did I work on Tuesday?" or have it build memory about your projects. The folder contains an AGENTS.md explaining the format to the model.

Thought it might be useful for someone so chucked it up on GitHub

⚡ Community Insights

Discussion Sentiment

80% Positive

Analyzed from 469 words in the discussion.

Discussion (16 Comments)Read Original on HackerNews

Dramatize•about 3 hours ago

One app that's using this technique (not exactly sure if it's the same) is Littlebird: https://littlebird.ai/

I also saw that HeyClicky started doing something similar but end up removing from the product.

jwr•about 1 hour ago

Hmm. What would be really helpful for me would be an enhancement of the TypeWhisper app for dictation that would be able to get the context of what I am dictating into and send it along with the prompt to an LLM.

Probably much simpler and much less of a privacy problem (I run my own local LLM for that purpose so that nothing leaves the machine).

zahlman•about 3 hours ago

> reads the text of your focused window every few seconds through the Accessibility API

> It writes plain markdown

Where are the formatting decisions coming from?

Dramatize•about 2 hours ago

So the markdown is scaffolding the app generates deterministically.

The ## heading is built from the block's timestamps, the app name and the window title. The frontmatter is per-day boilerplate. The file:/url: line is the window's backing document where the app exposes one. The captured text underneath is written exactly as the tree handed it over: plain lines, no reconstruction.

That lossiness is also why the file/URL references exist. Trying to rebuild a document's formatting from its accessibility tree is a losing game, so instead each block records where the real document lives, and the LLM reading the file can open the original if the fragments aren't enough. "Plain markdown" was meant as "a markdown file you can open anywhere", not "faithful markdown conversion of what you saw"

purplecats•about 5 hours ago

neat. why not screenshot and tesseract (videos/images/viewport/etc)

Dramatize•about 4 hours ago

Because you then have the macOS orange screen sharing warning/icon. I don't really want to record my screen, just the text is enough.

hetsaraiya•about 4 hours ago

Agree

latexr•about 1 hour ago

Project reasons aside, macOS has perfectly capable, local, native OCR that you can access programmatically, without having to install something else.

etchalon•about 4 hours ago

I experimented with this exact same approach earlier this year.

It's barely sufficient, because, bluntly, most apps just aren't wired up right.

So you end up having to hand code a lot of specific profiles for specific apps to make this work well, and even then, you don't quite get the right level of detail to make it work out.

Will try this app, to see if it improved on my own approach, but man, the hope levels are low.

albert_e•about 2 hours ago

I hope in near future _that_ layer of abstraction -- looking at a fairly standard application window with minor UI variations and reasoning about what area / labels within the UI mean what (possibly paired with app documentation) -- could probably become a light-weight fine-tuned vision model it itself that can run fully locally.

hetsaraiya•about 3 hours ago

I haven't read about how the Codex Appshots work yet, but this can be used to extract text properly. I guess. How does this idea look to you?

Dramatize•about 2 hours ago

Appshots is fine for what it is, but not great for building ambient context about your daily usage/work.

hetsaraiya•about 2 hours ago

But we can clean that extracted text from Appshots.

UPDATE: As i just checked it is also using Accessibility API. So i guess we will have access to the same set to data.

Dramatize•about 4 hours ago

Yeah don't get your hope up too much. I need to push through more cleaning up of what's captured. Let me know how it goes, keen for some feedback

zahlman•about 3 hours ago

One of the reasons people like TUIs is because the text is always just right there.

ShinyLeftPad•about 3 hours ago

Not if it's rendered on GPU, I guess?