PCS 0.9.114

PCS 0.9.114 - Connected memories, longer Local conversations, and experimental Linux & Mac builds

This release brings together the changes from 0.9.103–0.9.114: connected memories and lessons, a larger Continuity Map, more room for Local conversations, easier setup, expanded reading notes, and repairs across memory, voice, approvals and backups.

It also includes experimental Linux and Mac packages. If you have been waiting to try PCS on either platform, your feedback would be especially welcome.

Experimental Linux & Mac builds - testers wanted

These are early ports with more testing ahead of them. Installation reports, successful launches, problems and suggestions will all help improve them.

Mac - Apple Silicon

The first Mac experiment targets Apple Silicon Macs running macOS 13.3 or later, with native Python 3.12.

It includes experimental Metal and CPU paths for Local conversations, CPU speech processing and semantic search, and Mac file dialogs and launch scripts.

The application has not yet been run on a Mac. Automated checks and package inspection have been completed from another platform, but installation, Metal acceleration, microphone and speaker use, desktop integration and everyday reliability still need testing on actual Mac hardware.

This package is unsigned and not notarized. It uses setup and launch scripts rather than a standalone .app or DMG. Intel Macs and running through Rosetta are outside this first experiment.

The Mac version remains 0.9.114.

Linux - x86-64

The Linux experiment targets Ubuntu 24.04 on x86-64, with Python 3.12.

Limited native CPU checks have passed for Local inference, speech processing and generation, and semantic search. These used small, controlled workloads; they do not establish how the recommended larger models will perform.

The package also includes desktop launch scripts, file-dialog integration and an experimental Vulkan path for Local text generation. Graphical desktop use, physical microphone and speaker operation, GPU acceleration, OBS, suspend/resume and longer everyday sessions still need testing.

The Linux packages retain their 0.9.115 experimental version labels, although they are attached to this v0.9.114 release.

Getting started with the experiments

Both experimental downloads are source-based packages with setup and start scripts. They require Python and additional dependency downloads; optional models and voices remain separate downloads.

• Mac: read docs/MACOS_EXPERIMENTAL.md, then use Setup-PCS.command and Start-PCS.command.
• Linux: read docs/LINUX_EXPERIMENTAL.md, then use Setup-PCS.sh and Start-PCS.sh.

Please check the included requirements before choosing a Local model. The larger Windows CUDA context settings described below do not apply to these experimental backends.

Highlights

• Connect memories and preserve lessons. Open a memory and choose Settled, links & lessons to record how an idea developed, connect it to earlier experiences, or save a named lesson with its story and takeaway. Links refer to the exact revisions you reviewed. A settled mark applies to one revision and retires when that wording is corrected. Optional model suggestions remain proposals until accepted.

• Help conversations notice relevant memories. With Help notice relevant memories enabled, supported routes can use a compact overview of saved topics and lesson names. The overview is built locally and follows existing sharing choices; full stories and supporting records are retrieved separately.

• Explore a larger Continuity Map. Index the whole archive can include up to 50,000 memories, with an encrypted index that follows saves, corrections and forgetting. Maps support up to 250 themes, give quieter themes more room, and group connections into expandable lines.

• Longer Local conversations on Windows CUDA. The recommended NVIDIA CUDA profile now defaults to 128K context, with options up to 256K and FP16 or Q8_0 cache settings. Custom GGUF profiles retain their chosen window. Long Local sessions can compact their working context while keeping recent complete exchanges and making older wording searchable for that session. Saved originals remain unchanged.

• More time to finish speaking. Local microphone turns allow up to 60 seconds, with a warning at 50 seconds. The silence threshold increases from 0.8 to 1.5 seconds, giving you more room to pause during a thought.

• Less repeated Local work. Active replies take priority over eligible background memory formation. Unchanged prompt prefixes can be reused, interrupted model downloads can resume, and some speech or search cancellations avoid unnecessary worker restarts.

• Larger, resumable Prepared Notes. Local preparation can use portions of up to 80,000 source characters when the complete request fits. Prepare book saves progress after each approved portion, and Resume lets you review the next scope. Topic search can open relevant excerpts from saved notes; unreadable PDF ranges are recorded as omissions.

• Clearer setup and remembered approvals. Quick Setup brings connection and continuity choices together, including a conversation-only option. New-user Calendar assistance defaults to Only when I ask after review. Jev and proactive help can reuse approval for the exact model and input setup already reviewed; new combinations require their own review.

• More focused conversation and updated defaults. “Think with me” guidance encourages developing one manageable thought at a time. Vision’s default maximum image width increases to 1920 pixels, while preserving explicit saved choices. The Windows NVIDIA advisory warning for 16 GB-class cards moves to approximately 15.5 GiB.

Other repairs and improvements

Repairs cover spoken confirmations of memory corrections, clearer save feedback, stale prepared content after edits or forgetting, voice-recognition recovery, repeated Calendar requests, and longer cloud conversations.

Search improves matching for Chinese, Japanese, Korean, Indic and accented text. Materials, preparation status, Map actions, Stop and Lock handling, older Local configurations, passphrase changes and protected restores also receive fixes.

Automatic backup cleanup now removes only backups belonging to that installation. Retained links and lessons protect their still-available referenced revisions during ordinary history pruning; they cannot recover wording already pruned.

Before upgrading or trying an experimental build

Save your work, create an encrypted vault backup, and back up original Materials files separately. Quit PCS and extract the complete package into a fresh folder. Follow the included Backup & restore instructions.

Test experimental builds with a copy of your vault and keep an untouched backup. Saving native Mac or Linux settings can add compatibility requirements that stock Windows 0.9.114 cannot read—even though the Mac build has the same public version number.

For Windows upgrades, some features introduced since 0.9.102 raise the vault’s minimum reader requirement. Check backup reports the required reader. Windows 0.9.114 itself adds no further reader requirement beyond 0.9.113.

If you use proactive help, complete the updated review through Review model and inputs → Allow → Save settings.

Downloads

Platform	Application package	Matching source
Windows x64 - 0.9.114	Windows ZIP	Windows source ZIP
Linux x64 - 0.9.115 Experimental	Linux TAR.GZ	Linux source ZIP
Mac Apple Silicon - 0.9.114 Experimental 1	Mac ZIP	Mac source ZIP

Download SHA-256 checksums.

Choose the named package for your platform. Optional models, voices and runtime packs remain separate downloads where required.

Testing status

The included repair records describe automated source, interface and package checks. Linux has the limited native CPU checks described above; Mac execution remains unverified. Native Windows GPU, microphone, OBS and live-model acceptance also remain outstanding in the accompanying repair record.

These checks do not establish a general speed or recall-accuracy improvement.

Please share your feedback

If you try PCS—especially either experimental build—please open a GitHub issue and tell us how it went.

Useful details include:

• The package and version you downloaded.
• Your operating system version, CPU, GPU and RAM.
• Whether you used a cloud connection or Local, including the model and backend.
• What worked, what failed, and the steps needed to reproduce a problem.
• Relevant error messages, with private details removed.

Feedback on installation, first launch, microphone and speaker selection, Metal or Vulkan, file dialogs, OBS and suspend/resume would be particularly helpful. Reports that everything installed and started successfully are useful too.

Suggestions about confusing controls, missing features or how PCS fits into your everyday use are also welcome. Thank you for helping test these first steps onto Linux and Mac.

Support PCS

If PCS is useful to you, you can help support its development. Support is optional.