Today it flagged a post about an AI tool for HN and suggested I reply with:

"honestly, if you need an AI to sift through hn, you might be missing the point—this place is about the human touch. but hey, maybe it'll help some folks who just can't take the noise anymore."

So my AI, which I built specifically to sift through HN for me, is telling me to go flame someone else for doing that.

No deeper point here. I just thought it was really funny.

"honestly, if you need an AI to sift through hn, you might be missing the point—this place is about the human touch. but hey, maybe it'll help some folks who just can't take the noise anymore."

The really made me laugh - seems like AI is totally pushing me into hypocrisy haha

Did you think of building some proactive AI tools to make use of all this centralized data?

One thing I keep coming back to though - what if the tool could actually watch how you use your CRM and then suggest automations based on what it sees you doing repeatedly?

I've been building something called MemoryLane (https://github.com/deusXmachina-dev/memorylane) that does exactly this - it captures screen activity, spots repeated workflows, and suggests automations. Works as an MCP server so you can plug it into Claude or Cursor. Instead of you having to describe what you want automated, it just watches and proposes stuff.

Have you thought about adding something like pattern detection to denchclaw? Feels like it'd fit really well with the "everything app" direction. For us the most useful engine for executing skills and automations is surprisingly cowork thus far, haha

I've been working on a slightly different angle with MemoryLane (https://github.com/deusXmachina-dev/memorylane) - instead of giving agents a place to write their own memories, it captures the user's screen activity and makes it queryable. So the agent gets context about what the human was doing, not just what the agent itself did. It plugs in via MCP so Claude Code / Cursor can just ask it stuff.

I think there's something interesting in combining both - agent-vfs for the agent's own state, and something like MemoryLane for the human side. How do you think about that boundary between what the agent remembers vs what it knows about the user?

We need to use vector DBs just because of the amount of data. But on a different layer we want to help create file-based instrucations/skills for patterns that we detect and think can be automated.

I've been building MemoryLane (https://github.com/deusXmachina-dev/memorylane) which comes at this from the other side - it records screen activity, spots repeated patterns with AI, and then tells you "hey you keep doing this, want to automate it?" Works as an MCP plugin for Claude/Cursor.

Feels like pattern detection (finding what to automate) + browser agents like yours (actually doing the automation) is the right combo. Are you thinking about the discovery side at all, or mostly focused on execution?

We added a blacklist coordinator that checks four rules on every window-focus event:

- App name (1Password, Signal) — normalized across .app, .exe, bundle IDs

- Window title wildcards (*bank statement*)

- URL patterns (*://*.bank.com/*)

- Incognito auto-detection (on by default, zero config)

When anything matches, capture pauses and queued frames get flushed. Switch away, it resumes.

Surprising part: the matching logic was trivial. The annoying work was cross-platform app name normalization and making the settings UX simple enough that people actually configure it.

Comments URL: https://news.ycombinator.com/item?id=47309793

Points: 2

# Comments: 1

Processing hundreds of screenshots/hour forced us to optimize for token costs.

The surprise: send video, not images

- Single screenshot (1698×894): 1,812 tokens

- Same frame in video: 258 tokens (Gemini 2.5) or ~70 tokens (Gemini 3)

- Full 8-hour workday: ~$1-3

Video gives you timestamps for free and compresses well since consecutive frames are nearly identical. We keep costs down by having the LLM write short summaries while running OCR locally for text extraction.

Comments URL: https://news.ycombinator.com/item?id=47232766

Points: 3

# Comments: 1

Btw, I built something similar to solve the context problem for most of my laptop-based activity.

It's slightly more heavyweight (electron app ingesting screenshots) - that being said I took many similar design decisions (local embeddings, sqlite with vector search and FTS hybrid, MCP extension to claude). Feel free to check it out:

https://github.com/deusXmachina-dev/memorylane

Building an electron app (https://github.com/deusXmachina-dev/memorylane-private) which does A LOT of screenshots and this looks like something which could save the users (for now mostly us the dev team :D) a lot of money.

Different tradeoff - Total Recall is curated (daily logs) and lean, MemoryLane for now captures broadly and relies on search to surface what's relevant. I think they are complementary: your write gate keeps project knowledge tight, broad context with search fills in the "wait, what tab/webpage was that" gaps.

We applied to this batch (YC S26), still waiting on Apple notarization before a wider release. Happy to chat if anyone's curious about the screen-context approach.