AI

MineContext by ByteDance: A tiny desktop that remembers everything for you

What MineContext is, real-world use cases, why it could change knowledge work—and why it might be a secret weapon for ADHD brains

Field Report October 11, 2025
MineContext by ByteDance: A tiny desktop that remembers everything for you

If you’ve ever closed 27 tabs and immediately forgotten what problem you were solving, MineContext is the “external brain” you wish you had. It’s a proactive, context-aware desktop agent from ByteDance’s Volcengine team that quietly captures your on-screen work (screenshots + text understanding), organizes it, and then pushes you summaries, to-dos, prompts, and resurfaced references when they’re actually useful. Think “Spotlight search + personal historian + gentle project manager.”

MineContext concept art — your day becomes structured context blocks

What MineContext is (in one breath)

An open-source, local-first app (currently with a Mac download) that records your digital activity via screen monitoring, runs a processing pipeline (chunking, entity extraction, dedupe, embeddings), stores it in a local DB (SQLite/Chroma), and lets an LLM/VLM answer questions or proactively generate daily/weekly briefings, to-dos, and writing prompts from your real context. You bring your own API key (OpenAI or ByteDance’s Doubao, with Ollama planned).

Why that matters

  • Local-first privacy. Your captured context stays on device by design. For anything “lifeloggy,” that’s huge.
  • Proactive, not just reactive. It doesn’t wait for your perfect prompt; it nudges you with summaries and next steps.
  • Built like a platform. There’s a modular backend (FastAPI + managers for capture/processing/consumption) and a roadmap of context sources (files, MCP-enabled apps like Notion/Jira/Figma, RSS, email, even wearables).

How it works under the hood (nerd-friendly)

Capture → Process → Store → Resurface/Generate

  1. Capture: Periodic screenshots + (soon) files, links, meeting notes, browser extensions, MCP app data.
  2. Process: Chunking, entity extraction/normalization, multi-modal understanding, and embedding.
  3. Store: Local vector + relational storage (SQLite/Chroma) for fast similarity search.
  4. Consume:
    • Q&A over your day (“find that plot I made at 2:14 pm”).
    • Proactive daily/weekly summaries, to-do suggestions, and “you were working on…” nudges.

Today, the GUI download is for macOS; the backend is cross-plat Python (so Windows/Linux hackers can run services), but the polished desktop app target is Mac right now.


Concrete use cases I’m excited about

1) Writing & research (a.k.a. my life)

  • Auto-build a literature trail: papers, highlights, code snippets, screenshots of plots—then ask, “pull the three charts I used to argue X.”
  • Draft posts from your actual working context instead of a blank page.

2) Project management without project managing

  • It recognizes you bounced between a Jira ticket, a Figma spec, and a PR, then proposes a bullet update and a next-step checklist.
  • Weekly digest you can paste into a stand-up doc, no spelunking required.

3) Meeting memory

  • Capture slides and chat windows across calls; generate action items and a recap tied to screens you actually saw. (Roadmap includes meeting notes + MCP).

4) “Where did I see that?” recovery

  • Query by visual memory: “find the blue chart with CN0 on the y-axis I looked at after lunch.” The combo of screenshots + embeddings shines here.

How this could change the way we work

From “pull” to “push.” Most tools wait for your query. MineContext flips it: your workspace is passively indexed, and the system brings back relevant fragments when you resume a task. It’s the difference between rummaging and being handed the right tray.

From apps to contexts. The unit of work stops being “a file in App X.” It becomes a living context spanning docs, screenshots, chats, and tasks—resurfaced at the right time. That aligns with the wider shift toward agentic workflows where long-context models and MCP-style connectors orchestrate across tools.

From knowledge hoarding to knowledge circulation. Daily summaries + nudges reduce decay. Less “what was I doing?” and more “pick up where you left off.”


The ADHD angle: externalizing executive function

ADHD often challenges working memory, time management, and task switching—the executive-function layer that coordinates everything. External aids and structured cues measurably help.

Here’s how MineContext can help, practically:

  • Automatic thought capture. No friction to “brain-dump.” Screens become recallable context; you ask later, “what did I do in that 45-minute hyperfocus sprint?”
  • Context-aware nudges. Proactive to-dos and summaries create the external cues ADHD brains benefit from—timely prompts that reduce decision fatigue.
  • Reduce context switching tax. When you must switch, the agent preserves state and reconstructs it later—less re-onboarding each time.
  • Ritualize progress with weekly retros. Scheduled digests provide a body-double-ish accountability effect (“here’s what you shipped”), which the ADHD community often finds motivating.

Tip-sheet for ADHD-friendly setup (my picks):

  1. Short capture interval (e.g., 5–10s) for finer recall without manual notes.
  2. Daily + weekly briefs turned on; pipe the output into your task app.
  3. Create a “Restart Me” prompt the agent pins whenever you reopen a project: “Yesterday you touched A/B/C; next best action is D.”
  4. Pair it with a lightweight body-doubling session (Focusmate/FlowClub, or even a friend on video). Start the timer; let MineContext prep your kickoff checklist.

Privacy, ethics, and the “always watching” feeling

Lifelogging raises real privacy questions. MineContext’s local-first posture is a strong start, but you should still:

  • Define capture areas (not the whole screen).
  • Exclude personal windows (banking, 2FA).
  • Periodically purge + encrypt backups.
    The broader HCI literature flags privacy trade-offs in visual lifelogging; design your setup with that in mind.

Limitations today

  • Mac-first experience. Windows/Linux require tinkering with the backend; polished desktop builds aren’t the primary path (yet).
  • APIs & models. You’ll need your own LLM key (OpenAI/Doubao). Local models are on the roadmap, but quality/cost trade-offs apply.
  • Over-capture temptation. More data ≠ more clarity. Curate sources, or you’ll drown in screenshots. (Ask me how I know.)

Getting started (what I did)

  1. Grab the Mac app from the repo’s release page.
  2. Run the quarantine command the README suggests (macOS thing).
  3. Drop in your API key.
  4. Turn on Screen Monitor, pick a capture region, and… forget about it for a day. Let the agent surprise you with its first daily brief.

Bottom line

MineContext makes a persuasive case that the future of work isn’t five more apps—it’s a thin, local layer that continuously understands your context and helps you resume, recall, and ship. If your brain runs a little spicy (mine does), those timely nudges and “here’s-what-you-were-doing” breadcrumbs feel less like surveillance and more like kind structure.

My Amazon Picks

As an Amazon Associate I earn from qualifying purchases.

Shark AI Ultra Robot Vacuum with self-empty base

Shark AI Ultra Robot Vacuum

Matrix Clean charts every room, grabs the mess, then offloads dust into a 60-day HEPA self-empty base.

  • Hands-free scheduling and voice/app control with precise home mapping.
  • Anti-allergen filtration traps pet dander while the self-empty dock handles the bin.

Join the discussion

Thoughts, critiques, and curiosities are all welcome.