Self-hosted memory-as-a-service for AI agents.
v0.8.x: Memory redesigned as per-principal behavior with automatic extraction from conversations (AddMessages RPC, memory_messages worker). Per-topic vector backend configs, backend registry, Prometheus metrics, Helm hardening, Python and TypeScript SDKs, MCP server, CompactionService, chunk linking, managed pipeline, dashboard, creel-chat.
Creel provides topic-scoped memory with principal-based RBAC, server-side document processing, per-principal memory (Mem0-style), and dual-mode retrieval (semantic search with citations + temporal context). Upload documents and Creel handles extraction, chunking, and embedding automatically. It runs in your infrastructure via Helm chart.
# Start PostgreSQL + Creel server
docker compose up -d
# Build the CLI tools
make build
# Source the dev environment (sets CREEL_GRPC_ENDPOINT and CREEL_API_KEY)
source .env
# Create a topic and upload a document
bin/creel-cli topic create --slug my-notes --name "My Notes"
bin/creel-cli upload --topic my-notes --file notes.txt --name "Meeting Notes"
# Check processing status
bin/creel-cli jobs list --topic my-notes
# Search (once processing completes)
bin/creel-cli search --topic my-notes --query "action items" --top-k 5See Quickstart for the full walkthrough.
- Upload and forget: upload PDFs, HTML, or plain text; Creel extracts text, chunks it (fixed-size or LLM-based semantic), and embeds in the background
- Document citations: search results include document metadata (title, author, URL, date) for proper attribution
- Per-principal memory: automatic fact extraction from conversations with Mem0-style conflict resolution (ADD/UPDATE/DELETE/NOOP)
- Topic > Document > Chunk hierarchy with RBAC and cross-topic linking
- Dual-mode retrieval: RAG (semantic search) and context (temporal ordering)
- Server-driven compaction with link preservation
- Pluggable vector backends: pgvector (reference) with per-topic backend configs and a lazy-initialized backend registry
- 72 gRPC/REST RPCs across 10 services
- Two ingestion paths: managed (upload and forget) and direct (pre-chunked, pre-embedded) for power users
- Admin dashboard: Laravel-based web UI for config, topic, system account, and memory management
- creel-chat: interactive demo agent with streaming, memory, cross-topic RAG, and citation display
| Guide | Description |
|---|---|
| Quickstart | End-to-end setup in under 5 minutes |
| Fullstart | Hands-on walkthrough of every feature |
| Concepts | Data model, auth, search modes, memory, document processing |
| API Reference | All 63 RPCs with request/response details |
| Development | Dev environment, testing, adding RPCs and workers |
| Deployment | Helm chart, configuration, OIDC setup |
| creel-chat | Interactive demo agent with conversation memory |
| Architecture | Full design document and roadmap |
Active development (v0.8.x). Phases 1-10 complete: core storage, RAG, auth, managed pipeline, per-principal memory with automatic extraction, compaction, integration layers (SDKs, MCP, tool schemas), and backend hardening (per-topic vector backends, Prometheus metrics, Helm production readiness). See docs/ARCHITECTURE.md for the roadmap and CHANGELOG.md for release history.
Apache 2.0. See LICENSE.