A Bonfire is the knowledge engine. It's the persistent store that owns all of a community's accumulated intelligence.
## What a Bonfire Owns
**Documents and chunks.** All ingested content — uploaded documents, transcripts, raw text — is split into chunks and stored. These are the raw material for everything downstream.
**Knowledge graph.** A Neo4j graph (via Graphiti) containing entities, relationships, and episodes extracted from conversations and documents. This is the structured representation of what your community knows.
**Vector store.** Semantic embeddings (Weaviate) of all chunks and labels. This enables search by meaning — find content similar to a query rather than matching keywords.
**Taxonomy and labels.** Auto-generated multi-label classification of your content. Labels are applied at the summary level and propagated to chunks, driving content organization and discovery.
## Bonfire ≠ Agent
[[files/Technical/Agents|Agents]] are stateless interfaces. They don't store anything. They read from and write to the Bonfire.
One Bonfire can have many agents. All agents in a Bonfire share the same knowledge graph, documents, and vector store. Different windows into the same knowledge base.
```
┌──────────────────────────┐
│ Bonfire │
│ │
│ Documents · Graph │
│ Vectors · Taxonomy │
└─────────┬────────────────┘
│
┌──────────┼──────────┐
│ │ │
Agent A Agent B Agent C
(Telegram) (Discord) (API)
```
## How Knowledge Flows In
**Conversation capture (automatic):** Every 20 minutes, a background process takes recent messages from each agent's stack, extracts entities and relationships, creates episodic summaries, and writes everything to the Bonfire's knowledge graph. This happens for all messages — including ones the agent didn't respond to.
**Document ingestion (manual):** Upload PDFs, text files, or raw content via API. Content is chunked, summarized, labeled with taxonomy categories, and embedded into the vector store.
**Episode updates (API):** External systems can write episodes directly to the knowledge graph.
## The Content Pipeline
```
Content arrives
│
▼
Split into chunks → stored in MongoDB
│
▼
Generate summaries (async job)
│
▼
Generate taxonomy → label chunks (async jobs)
│
▼
Embed into vector store (Weaviate)
│
▼
Extract entities & relationships → knowledge graph (Neo4j)
│
▼
Content is searchable via:
• Semantic search (vector store)
• Knowledge graph search (Delve)
• Agent chat
```
## Searching a Bonfire
**Delve search** — Unified semantic search across the knowledge graph. Returns entities, relationships, and episodes ranked by relevance.
**Vector search** — Find document chunks or labels semantically similar to a query.
**Agent chat** — Ask the agent in your group chat. It combines knowledge graph results, vector search results, and recent conversation context before responding.
**Graph explorer** — Visual exploration at [graph.bonfires.ai](https://graph.bonfires.ai).
**MCP** — Connect Claude Desktop, Cursor, or other MCP-compatible tools to query your Bonfire programmatically.
## DataRooms and HyperBlogs
DataRooms package a Bonfire's content for the marketplace. [[files/Integrations & Use Cases/HyperBlogs|HyperBlogs]] are AI-generated articles created from DataRoom knowledge, purchasable onchain. See [[HyperBlogs (Technical)]] for the full architecture.
---
**See also:** [[files/Technical/Agents|Agents]] · [[docs/docs26/kEngrams]] · [[HyperBlogs (Technical)]] · [[docs/Introduction]]