Drivesoid

An emotional drive system for AI personas. Tracks 16 psychological dimensions (attachment, threat, reward, etc.) that evolve in real time based on conversations, sleep cycles, and calendar events.

How it works

Drivesoid runs as a lightweight HTTP sidecar alongside your AI bridge. Your bridge reports conversation events; Drivesoid maintains an emotional state that your AI can read and incorporate into its responses.

Dimensions tracked: vitality, fatigue, longing, intimacy, possessiveness, lust, jealousy, anxiety, protectiveness, fear, contentment, elation, seeking, play, dejection, irritability

Quick start

For AI agents: Read AGENT_SETUP.md — it tells you exactly what to install, what to ask the user, and how to integrate.

For humans: Start the service — setup runs automatically on first launch:

npm start

Then open http://127.0.0.1:3001/setup in your browser and fill in the form.

Requirements

Node.js ≥ 18
Python 3 (for Claude Code hooks)
An OpenAI-compatible API key (DeepSeek recommended: deepseek-v4-flash)

Configuration

Copy drives.config.example.json → drives.config.json and fill in:

Field	Description
`persona.name`	Your AI's name
`user.name`	Your name
`relation`	`romantic` / `companion` / `friend`
`timezone_offset_hours`	Local timezone offset (e.g. `8` for UTC+8)
`classifier.endpoint`	API base URL
`classifier.model`	Model name (required, e.g. `deepseek-v4-flash` or `gpt-4o-mini`)
`classifier.api_key_env`	Env var name holding the API key
`server.port`	Port to listen on (default: `3001`)

Set your API key in .env:

DRIVES_API_KEY=sk-...

API

Endpoint	Method	Description
`/api/drives/status`	GET	Current emotional state (JSON)
`/api/drives/context`	GET	Pre-formatted `[drives]` block for system prompt injection (plain text; empty when stale)
`/dashboard`	GET	Live dashboard — view state, tune parameters
`/internal/drives/event`	POST	Ingest a conversation event
`/internal/drives/session-start`	POST	Catch-up tick at session start
`/internal/drives/sleep`	POST	Sleep start/end events

/internal/* accepts loopback connections only (127.0.0.1). /api/* is unrestricted.

Classifier labels

Each user message is classified into one of these labels, which drive dimension changes:

Label	When it applies
`affectionate`	warm, loving, expressing care
`playful`	teasing, joking, playful banter
`vulnerable`	expressing vulnerability or emotional fragility
`reassuring`	comforting, affirming, offering support
`intimate_reference`	referencing physical intimacy or the body
`intimate_event`	actively engaged in an intimate interaction
`struggling`	stress, exhaustion, feeling unable to cope
`cold`	emotionally withdrawn, detached, terse
`distant`	distracted, disengaged, doesn't want to talk
`conflict`	genuine mutual argument, emotional escalation
`hostile`	one-sided attack or harsh words
`fear_separation`	fear of the other person leaving or being absent
`fear_death`	fear of death, suicide, or life-threatening situations
`fear_concern`	worrying about something bad happening to the other person
`fear_general`	other fears not covered above
`neutral`	ordinary response, no notable emotional tone

Reset

Wipe config and data to start fresh:

npm run reset
npm start

License

MIT