An AI layer that reads an online community, remembers who genuinely contributes and how the group is changing, and helps you moderate it and reward the people who carry it.
Communities live on Telegram and Discord, but what happens inside them is invisible and ungoverned. Moderators can't tell genuine contributors from farmers, bots, and noise; volume gets rewarded over value; a trusted member turning toxic or a coordinated farm goes unnoticed until it's too late. Keyword filters and ban-bots don't understand context, and they have no memory.
Cotribe installs as a bot and changes that.
- Reads every message in context — against each community's own norms, not a keyword list. Genuine help, real insight, and welcoming newcomers count; disagreement and criticism are free; only toxicity, spam, farming, and scams are penalized.
- Builds a farm-resistant reputation — a trust graph from who genuinely endorses whom, weighted so sockpuppet rings and reply-bribery can't inflate themselves.
- Remembers every member and the group over time — a living memory of standing, behavior, and sentiment, so Cotribe catches change: a member turning toxic, a farm account, a rising contributor, a community cooling off.
- Assists moderation — surfaces who needs attention and shows the whole community's health at a glance.
- Rewards the real contributors — turns reputation into merit-based rewards for the people who actually build the community.
Four layers — raw activity in, rewards out:
1 · Data. Ingests every message, member, and interaction into a durable, platform-neutral record. Message text is never stored downstream or shown; an interaction is "A replied to B," never what was said.
2 · Analyze. An LLM scores genuine contribution — not volume — against the community's norms, while a trust graph and an authenticity model resist farming and sockpuppets. Output: per-member contribution, trust, authenticity, and roles.
3 · Memory. The analysis becomes a living memory in Cotribe's own store: each member's reputation and what they're known for, plus the group's sentiment and trajectory over time. It's queryable — "who really helps here?", "how is this member doing?"
4 · Reward. Reputation becomes merit-based rewards. Genuine contributors earn offchain XP proportional to their real contribution and the trust of the peers who endorse them; farm, toxic, and scam accounts earn nothing, however much they post.
An AI agent sits across the Memory layer and makes the calls. It understands context and — crucially — remembers members across sessions, so it can moderate a trajectory: it knows someone was helpful last week and is turning hostile now. The memory lives in Cotribe's store; the Mind is the brain that reads it and acts — swappable, but central.
A real-time bubble graph of the community — members as bubbles (size = standing), interactions as lines (green = endorsing, red = conflict), colored by health. A moderator sees clusters, isolated bad actors, the XP leaderboard, and the community's overall health at a glance.
You need a Telegram bot (added to your group, privacy mode off) and a Mind from
hellominds.ai. Put them in a .env at the repo root:
TELEGRAM_BOT_TOKEN=...
MINDS_BUILDER_API_KEY=...
COTRIBE_CONCIERGE_MIND_ID=...
pnpm install
# See your real group as a live bubble graph
pnpm demo:live # → http://localhost:5055, then chat in your group
# Watch the Mind read a community and remember it across sessions
pnpm demo:mind
# Reward genuine contributors with offchain XP — merit, not volume
pnpm demo:rewardsrc/ is the full analysis pipeline — ingest, LLM scoring, the trust graph,
authenticity, roles, and the memory layer with its swappable agent (an offline
deterministic reader, an LLM, or a Mind). It runs on Postgres and Redis and exposes
the live graph as a web view. The demos above talk to the Telegram and Mind APIs
directly, so you can see Cotribe work without standing up the full stack.
Cotribe works on derived signals — counts, trust, roles, memory — never on stored raw message text, and it never displays what anyone said.