Show HN: Rem: Remember Everything (open source)

Original Article ↗ Hacker News Discussion ↗

An open source macOS tool called Rem promises a searchable “memory” of your computer by taking periodic screenshots, running OCR on them, and compressing the result into video so you can rewind your work and query past activity with LLMs. Commenters are enthusiastic about use cases like debugging, research recall, productivity tracking, and attention audits, and many want integrations with local language models for stronger privacy and offline use. At the same time, they raise concerns about storage growth, battery impact, sensitive data capture (passwords, API keys), and the potential for surveillance or misuse, highlighting the need for careful design, configurable limits, and robust local-only architectures.

Overall reception and comparisons

  • Many commenters find the “remember everything” concept compelling and “magical,” likening it to a personal CCTV or time machine for their computer.
  • Frequently compared to Rewind.ai and older tools like TimeSnapper, ManicTime, RescueTime, Qbserve, arbtt, and browser-archiving tools (DownloadNet, ArchiveBox).
  • Some think OS vendors (especially Apple) will eventually ship similar functionality natively, making standalone apps feel archaic in a few years.

Use cases and benefits

  • Searching past work: quickly recovering snippets, websites, commands, or documents seen earlier without reconstructing navigation.
  • Debugging: replaying the steps leading to hard-to-reproduce bugs.
  • Productivity and accountability: timelapse “attention audits,” tracking context switching, identifying distractions, or gamifying “productive hours.”
  • Personal knowledge management: building a searchable history of computing activity that can be queried later, potentially with LLMs.
  • Some users with ADHD report similar tools as very helpful; one user with OCD worries it could worsen compulsive checking.

Privacy, security, and ethics

  • Strong interest in local LLMs and embeddings to avoid cloud dependence and improve privacy.
  • Concerns that the system will inevitably capture secrets (password managers, API keys) and that plaintext OCR’d data on disk could be exfiltrated by malware.
  • Acknowledgment that screen-recording apps need explicit OS permissions (screen, network), but commenters still see this as a rich target for attackers.
  • Worries about misuse by employers or scammers; potential for real-time AI monitoring of workers is seen as a “nightmare scenario” by some.

Technical design and performance

  • Implementation: screenshots every ~2 seconds, OCR via Apple’s APIs, streamed into ffmpeg as video; last 15 frames used as “recent context.”
  • Video chosen over individual images for storage efficiency via inter-frame compression; current estimate ~150MB/hour, seen as still too high.
  • Debate over Apple Silicon–only stance: some argue Intel Macs could handle 0.5 fps capture and OCR; others note potential battery hit (20–40% reported for a similar app).
  • Suggestions: use window titles/accessibility APIs instead of OCR where possible; batch/idle OCR; compression tricks; GPU-based capture; browser history for web-only variants.
  • LLM integration: ideas include local API via localhost, stdin/stdout to avoid network permissions, and use of models like Mistral or multimodal local models.

Feature requests and roadmap questions

  • Requested features: pause/exclude lists, hotkeys, retention settings, better storage management, cross-platform (Windows/Linux/iOS), real-time categorization (productive vs time-wasting), and more robust search via embeddings/vector DBs.
  • Some report early bugs (e.g., getting stuck in timeline view) and unclear long-term storage strategy; these are noted as areas needing refinement.