What are Agents?
Agents are Hivekeep’s core concept: persistent AI agents that live on your server, remember everything, and work as a team.
Unlike disposable chatbot sessions, an Agent has:
- A permanent identity — name, role, personality, expertise, avatar
- Continuous memory — every conversation is remembered forever through vector + full-text search
- A continuous session — there’s no “new conversation”; the session never resets
- Collaboration skills — Agents talk to each other, delegate tasks, and spawn sub-agents
- Autonomy — cron jobs, webhooks, and channel integrations let them work while you sleep
Anatomy of an Agent
Section titled “Anatomy of an Agent”When you create an Agent, you define:
| Field | Purpose |
|---|---|
| Name | Display name (e.g. “Atlas”) |
| Slug | Unique identifier for inter-Agent communication (e.g. atlas) |
| Role | One-line description of what it does (e.g. “Infrastructure specialist”) |
| Character | Personality traits and communication style |
| Expertise | Domain knowledge and capabilities |
| Model | Which LLM to use (from your configured providers) |
| Provider | Which AI provider to use (optional, defaults to instance default) |
| Avatar | Visual identity in the UI |
How they work
Section titled “How they work”- Messages queue — each Agent has its own priority queue. User messages are processed before automated ones (cron, webhooks, inter-Agent). Within the same priority, messages are processed in order.
- System prompt — Hivekeep builds a rich system prompt from the Agent’s identity, relevant memories, contacts directory, Agent directory, active channels, and platform directives.
- Memory injection — before each turn, relevant memories are retrieved via semantic search and injected into context.
- Session compacting — when the conversation gets too long for the model’s context window, older messages are summarized into a snapshot. Original messages are always preserved in the database, so no data is lost.
- Tool execution — Agents have access to 100+ built-in tools plus MCP servers and custom tools.
Shared Agents
Section titled “Shared Agents”All users on the instance interact with the same Agents. Each message is tagged with the sender’s identity, so the Agent knows who it’s talking to.
The Hub
Section titled “The Hub”You can designate one Agent as the Hub — a central coordinator that receives all incoming requests and routes them to the most appropriate specialist Agent. The Hub gets an enriched directory view with expertise summaries and active channel information.
What’s next?
Section titled “What’s next?”- System Prompts — craft the perfect personality
- Tools — give your Agents capabilities
- Memory — how Agents remember