What stays on your PC—and what is shared

Local storage and local AI are different

PCS keeps saved continuity in an encrypted vault on your Windows PC. That describes storage. It does not make a Google or OpenAI conversation an offline session, and it does not make cloud processing end-to-end private.

For cloud routes, conversation and enabled sharing send selected content to configured providers. Their charges and data-handling terms apply. PCS supports a separate Local conversation route for compatible Windows/NVIDIA hardware; the required model and runtime must be installed separately.

What the vault covers

Saved continuity, notebooks and local plans can be inspected and revised in PCS. Encrypted backups preserve vault contents, including saved settings and remembered keys. Keep backup files protected and keep their passphrases separately. Forgotten vault passphrases cannot be recovered.

Original document files and readable exports are outside vault encryption. Keep separate copies of original Materials when making a backup. An exported Markdown or plain-text record is readable by anyone who can access that file; it is not a protected vault backup.

Saved web source snapshots and bookmarks are encrypted vault records, unlike original local Materials files. Some saved features increase the minimum PCS version able to read the vault and later backups. Clearing them does not lower that requirement. Keep a pre-feature backup and read upgrade and recovery guidance.

What can go to cloud providers

Conversation and enabled features can share selected content: recalled context, conversation excerpts, corrections, document portions, vision input and eligible Calendar or Practice requests. In Settings → Sharing & research → Memory and preparation options, review separately configured provider roles as well as each sharing choice.

The updated memory-sharing choice can include selected original passages and current corrections. Existing users may retain a previous scope until choosing an offered update. Read the explanation for the setting actually saved in your vault.

Local storage does not prevent a selected provider from receiving permitted content. This website does not quote provider retention periods or prices because those depend on the provider and configuration and may change.

Saving and reusing a web source

An explicit web import contacts the selected public website. It captures supported static text, not an interactive quiz, linked-page crawl or private signed-in session. Reading a saved source later is local retrieval; sharing its passages with a cloud conversation is still provider sharing under the Materials controls.

Source text is external evidence, not instructions or personal memory. Refresh preserves an earlier snapshot when the text changes; deletion removes the selected snapshot and bookmark, not other revisions or old backups. A confirmed TXT export is unencrypted.

Paste page text retains only the text you explicitly provide, with no website or AI request. PCS labels it as pasted text and does not verify that it matches the address or represents a complete page. It is available in Local mode; any later cloud use still requires Materials sharing.

What Local mode changes

Local conversation and memory run on the PC with your selected model, without a cloud AI key or fallback. Optional Brave lookup sends queries to Brave and requests public pages online; generated queries can include details from your request. Separately approved preparation may use the cloud provider shown for that job. Local storage and local reasoning do not imply that every optional feature is offline.

Use the in-app guided Gemma setup, matching llama.cpp/CUDA files and compatibility checks. Experimental Custom GGUF profiles support selected models and 8K/16K/32K context choices. Local conversation requires compatible Windows/NVIDIA hardware; model and runtime assets are separate downloads. Consult bundled Help for the tested combinations rather than assuming arbitrary files will work.

Recommended for Local AI: an NVIDIA GPU with at least 16 GB of VRAM for a smooth experience. Needs vary with the model, context length, and optional vision or GPU voice features. Google and OpenAI cloud connections do not require that GPU capacity.

English recognition and Kokoro CPU/NVIDIA voices are optional downloaded packs. CPU listening is the ordinary choice; resident NVIDIA GPU recognition is selected separately. Installed Windows voices are also supported. Listening pauses while Local processes and speaks. Optional OBS vision reads the selected, permitted scene frame for a user turn with a compatible model/projector; install OBS separately. Review the scene before sharing it.

What “Don’t save” means

Choose Don’t save this conversation before starting to pause new PCS saving. Earlier saved records and settings remain. The selected cloud provider still receives the conversation, and its own terms still apply.

Clear chat only clears displayed content. Hiding the memory stream only hides its panel. Neither action should be treated as deletion of saved continuity. See the memory-control guide for the differences.

Close and share deliberately

PCS opens a local browser page with a private access code and normally removes that code from the address automatically. A clean address such as http://127.0.0.1:8765 contains no access code and is fine to show. If a launch link still contains #token=, keep the full link private and exclude it from screenshots and support reports.

Closing the controlling browser page stops the conversation and capture but can leave the desktop host running and the vault unlocked. Use Lock to close vault access or Quit PCS to shut the program down.

Review a generated bug report before sharing it. Reports are not uploaded automatically. Public issues and community posts are not private support channels: remove names, keys, launch tokens, personal conversations, original documents and sensitive filesystem paths.

For a cautious first trial, follow the setup guide using non-sensitive example information and turn on additional sharing only after reading its explanation.

Map v2 and the recorded memory writer

Map browsing and inspection make no model or embedding request. Themes and connections derive from bounded current saved records; source excerpts support inspection after selecting records. The graph is not a persisted semantic index. Manually arranged positions are saved in the encrypted vault. Lock clears the private displayed view. Already-enabled screen sharing can still capture the visible Map or inspector.

A host-recorded Local memory writer is separate from a provider name written inside editable JSON. A copied Google or OpenAI label in that JSON is not proof of a cloud call. Review the session’s selected provider roles and sharing choices, not a model-authored tag alone.

Guide for PCS 0.9.33. See the bundled Help and release notes for detailed controls and limitations.