Sity — AI assistant with persistent memory and its own cognitive architecture

Description and use

Sity is a self‑hosted conversational AI running on a Raspberry Pi 4B, combining local TTS/STT and SQLite memory with the Claude Haiku API, featuring a full cognitive pipeline and proactive behavior.

Problem

Developing a self‑hosted conversational AI that can remember past interactions, track user expectations, and take initiative, while running entirely on a Raspberry Pi 4B.

Outcome

Sity demonstrates that a fully local conversational AI can be built on modest hardware, providing persistent memory and proactive behavior while delegating only language generation to an external API.

Workflow

• 1. Capture user input via faster‑whisper STT on the Pi
• 2. Process input through perception module
• 3. Appraise emotional content
• 4. Update goal tracker and episodic/semantic memory in SQLite
• 5. Evaluate 10 possible actions using a utility‑based engine
• 6. Generate response with Claude Haiku API
• 7. Convert response to speech with Piper TTS
• 8. Reflect and update self‑model

Evidence

It's a self‑hosted conversational AI that runs entirely on a Raspberry Pi 4B (backend, database, TTS/STT — all local).

Source

The language model is Claude Haiku via API, but everything else is on‑device.

Source

Each conversation turn goes through: Perception → emotional appraisal → goal tracking Episodic and semantic memory that persists across sessions A utility-based decision engine (10 possible actions) Reflection and self-model that update over time

Source

Runs on Pi 4B (4GB RAM) SQLite database, no external dependencies Piper TTS + faster-whisper STT (both local) Caddy reverse proxy + Cloudflare Tunnel for remote access AGPL-3.0

Source