Skip to content

Familiar-Connect

An AI "familiar" that joins Discord voice channels, listens to users, understands speech, and talks back using real AI voices.

What it does today

  • Joins Discord text and voice channels via per-channel subscription commands (/subscribe-text, /subscribe-my-voice).
  • Runs every reply through a context pipeline of pluggable providers (character card, recent history + rolling summary, agentic memory search) and processors (stepped-thinking pre-processor, recast post-processor).
  • Replies via OpenRouter (main model + a cheap side-model slot for sub-tasks) and Cartesia TTS for voice playback.
  • Watches Twitch channels via EventSub and feeds events (subs, bits, cheers, follows, ad breaks, channel-point redemptions) into the same conversation pipeline.
  • Stores every transcript verbatim in SQLite; stores every familiar's long-term memory as plain Markdown files under data/familiars/<id>/memory/.

Where to look next

Status at a glance

Area Status
Context pipeline, providers, processors, budgeter Shipped
Memory directory + MemoryStore + ContentSearchProvider Shipped
Character card unpacker, SillyTavern lorebook importer Shipped
Per-channel subscriptions + channel modes + configuration model Shipped
Twitch EventSub integration Shipped
Metrics and profiling (per-turn traces + CLI report) Shipped
Conversation monitor (chattiness / interjection / lull) Shipped
Voice speech-to-text wired into the reply path Shipped
Typing simulation for text-RP channels Shipped
Post-session memory writer Shipped
Discord-side voice channel logging (threads + live embeds) Planned
Web search as a provider tool Research