Your AI’s memory, now with actual synapses

What it does Polywise bundles a local server, a desktop Electron app, and a CLI around SQLite and vector search to create a self-hosted AI workspace. You chat with LLMs through sessions, stash web clippings and notes, and promote recurring prompt patterns into reusable agents, provided you wire up at least one model provider. Everything stays local by default, retrieval included.

The interesting bit The authors cite 1940s–70s neuroscience—Hebbian learning and spreading-activation theory—as the blueprint for how knowledge should organically strengthen, connect, and even strategically fade. It is unusual to see a practical AI tool ground its memory model in synaptic potentiation papers rather than vector-database marketing.

Key highlights

• Runs fully self-hosted with an optional Electron desktop app and CLI wrapper.
• Uses SQLite with sqlite-vec for on-device vector search and retrieval, so your data never needs to leave the box.
• Organizes work into four concrete buckets: Session for chat, Linkcase for web extraction, Agent for reusable prompt templates, and Posts for persistent knowledge.
• Supports local LLM inference through node-llama-cpp in addition to remote providers.
• @ and / shortcuts let you pull files, agents, and tools into an active session without leaving the keyboard.

Caveats

• The README pitches a biological memory model—connections that strengthen and strategically fade—but never explains how those mechanics are implemented over a standard SQLite and vector-search stack.
• The embedded intro video relies on a local file path that will not render for anyone browsing the repository on the web.

Verdict Worth a look if you want a local-first, SQLite-backed AI workspace and are happy to supply your own model providers. Skip it if you need a managed SaaS or a plug-and-play solution that abstracts away all backend setup.