Media
2 items
Description
Tell your AI once — all your AI assistants remember too. One private memory layer across your AIs, kept locally on your device. Tell an AI assistant something on Monday, and your other AI tools know it on Tuesday. sovseal gives every AI you use one private, shared memory — and unlike tools that send your raw chat histories to external APIs or third-party servers, your memories never leave your machine. The memory engine — semantic embeddings, vector search, and AES-256-GCM encryption — runs entirely inside the extension. The embedding model ships inside the package, so capture and recall work offline, on a fresh install, with no account and nothing to set up. FEATURES • Install-and-Go — The embedding and vector engine runs inside the extension, on your device. No terminal install, no account, no server, and no first-run download: the embedding model is bundled in the package. • Broad AI Compatibility — Works seamlessly across major web-based AI chat interfaces. • Truly Local — All chat capture, vector indexing, and semantic recall happen on your device. Memory text is encrypted at rest with AES-256-GCM under a key that never leaves your browser. • Works Offline — Because the model is bundled, storing and recalling memories succeed with the network fully disabled. • Transparent First-Run Disclosure — A plain-language onboarding screen on install explains on-device storage, capture controls, and what the extension does and does not send. • Intelligent Capture-on-Send — Prompts are captured as you submit them, with a global pause switch and per-site controls. • Silent Automatic Recall — Relevant memories are injected into your outgoing prompt at the network layer, keeping the chat interface clean with no copy-pasting and no extra clicks. • In-Page Overlay — Press ⌘/Ctrl+M or click the launcher widget to review recent memories and delete individual entries. • Optional Zero-Knowledge Sync — Sign in to sync across devices. Blocks that leave your device carry ciphertext only; embedding vectors never sync at all. Sign-in is never required. • Experimental On-Device Fact Extraction — Off by default. If enabled, an open-source model downloads once on demand (from 283 MB), is SHA-256-verified before use, and runs locally on WebGPU. HOW TO USE 1. Install the extension — a first-run onboarding screen opens automatically explaining your privacy choices. 2. Click the toolbar icon to confirm the "On-device engine active" status. 3. Open any supported web-based AI chat interface. 4. Send messages normally. The local engine captures context automatically and builds an encrypted index on your device. 5. In later conversations the extension silently retrieves and injects relevant memories. Press ⌘/Ctrl+M or click the launcher to review or delete entries. PRIVACY Your privacy is the engineering principle this product is built on. The extension sends nothing to sovseal — there are no sovseal analytics, accounts, or servers in the capture and recall path. All embedding generation and vector indexing run locally on your device, and the embedding model is bundled inside the package, so capture and recall operate fully offline with zero network requests. If you explicitly enable the optional, experimental "Extract clean facts" setting, an open-source model is downloaded once on demand from Hugging Face, integrity-verified against a pinned SHA-256 digest before it is loaded, and executed locally on your GPU. Your conversation text is never sent to Hugging Face or to any external server. One honest detail about automatic recall: the memories sovseal recalls are appended to the outgoing message sent to your chosen AI provider. That recalled text reaches that provider as part of your active prompt, under their terms, exactly as if you had typed it yourself. This augmentation happens entirely in your browser. You can pause capture globally or per site, and delete any record, at any time. SUPPORT Found a bug or have a feature request? Open an issue at https://github.com/sovseal/core or email ops@sovseal.com. Version 0.4.1