sovseal — Private AI Memory● live

Give Claude, Cursor, and every MCP client one shared memory that never leaves your machine. On-device vector search via LanceDB + Transformers.js (384-dim embeddings), AES-256-GCM client-side encryption, and write-behind ciphertext sync. ### What you can do 🧠 **Store memory** — embed and persist any context fact on-device. Automatic de-duplication, PII redaction, and write-behind encrypted sync. 🔍 **Recall memory** — semantic search over stored context. 10 ms warm recall, 0 RTT reads. Network is never on the read path. 📋 **Briefing context** — auto-summarized, reinforcement-aware briefing of procedural and semantic memories for bootstrapping new conversations. ### Tools - `store_memory` — embed content on-device (384-dim ONNX) → write local LanceDB → return. Ciphertext sync runs write-behind; nothing blocks. - `recall_memory` — embed query (LRU-cached) → vector search local LanceDB → top-K by L2 distance. 0 RTT. ### Resources - `sovseal://context/briefing` — summarized, reinforcement-aware briefing for conversation bootstrap ### How it stays private - **AES-256-GCM** client-side encryption with 96-bit random IV per snapshot - 256-bit key stays in `~/.sovseal/config.json` (mode 0600) — server only sees ciphertext - **Verified Semantic Recall (VSR)** — every load re-derives SHA-256 and fails closed on mismatch - **Packet-Capture Guarantee** — run Wireshark against the server; if you find unencrypted context on the wire, the software is free forever ### Performance | Workload | Operation | p50 | p95 | | --- | --- | --- | --- | | 10K records · 1K queries | recall_memory (warm) | 6.1 ms | 10.4 ms | | Single write | store_memory | 3.8 ms | 7.2 ms | **No API key required.** Zero cloud dependency for local operations. Replaces: Mem0 • Zep • custom SQLite memory • in-context window stuffing