v1.0 — Now with Memory Governance

Memory that flows.

Knowledge that stays.

Self-hosted RAG infrastructure that gives your AI coding assistants persistent memory, semantic search, and knowledge governance.

Deploy in 5 Minutes View on GitHub

Free and open source. MIT Licensed.

reka — index session
$ reka index --project shared-ai-infra ► Scanning project structure...   Found 47 files across 6 directories ► Generating embeddings (BGE-M3)...   ████████████████████████░░░░  82% (39/47) ► Indexing to Qdrant vector store...   Collections: shared-ai-infra_codebase ✓               shared-ai-infra_docs ✓               shared-ai-infra_memory ✓ ► Building knowledge graph...   Nodes: 234  Edges: 891  Clusters: 12 ✓ Index complete in 4.2s — ready to query
The Problem

Your AI assistant forgets everything. Every. Single. Time.

Groundhog Day Debugging

You explain the same architectural decisions every session. Context resets, knowledge resets.

Codebase Blindness

Your assistant ignores existing patterns and duplicates utilities already built.

Vendor Lock-in & Data Leakage

Cloud memory means your code flows through third-party servers.

How It Works

From zero to persistent memory in four steps

1

Deploy

One docker-compose up command. Under 60 seconds.

2

Connect

Point your AI assistant via MCP. One JSON config.

3

Index

Crawls codebase, extracts symbols, builds dependency graph.

4

Remember

Persistent memory. Decisions carry forward across sessions.

Features

Everything your AI assistant needs to understand your codebase

Semantic Code Search

Hybrid vector + keyword search. Find code by what it does, not just what it's named.

Dependency Graph

Maps imports and relationships. Trace blast radius of any change.

Memory Governance

New

Auto memories enter quarantine. Human-verified become durable knowledge.

Symbol Index

Instant function, class, and type lookup by name across your entire codebase.

MCP Native

35 tools for Claude Code, Cursor, Windsurf. Zero configuration overhead.

Multi-Project Isolation

Namespaced collections per project. Share the server, never the data.

Memory Governance

Not all knowledge deserves permanence

Reka introduces a governance layer inspired by how human memory actually works.

Quarantine Zone

Every auto-generated memory starts here with a confidence score.

Review Pipeline

Contradiction detection flags conflicts with existing knowledge.

Durable Knowledge

Canonical knowledge base — decisions, patterns, insights.

memory_entry.json
{
  "type": "architectural_decision",
  "status": "durable",
  "confidence": 0.94,
  "content": "Use BullMQ for async...",
  "verified_by": "human_review"
}
Architecture

Transparent by design. No black boxes.

Claude Code
Cursor
Any MCP Client
MCP Protocol
MCP Server 35 tools · project isolation
HTTP / X-Project-Name header
RAG API Express :3100 · TypeScript
Qdrant Vectors :6333
Ollama / LLM Local inference :11434
BGE-M3 Embeddings :8080
Redis Cache :6380
Comparison

How Reka compares

Feature Reka Mem0 Zep Custom RAG
Self-hosted
Memory governance
MCP native
Dependency graph
Semantic search
Code-aware chunking
Setup time
5 min
30 min
15 min
Days
Cost
Free
$
$$
Time
QUICK START

Up and running in three commands

# Clone and start Reka git clone https://github.com/reka-ai/reka.git cd reka && docker-compose up -d   # Configure your MCP client cat > .mcp.json << 'EOF' {   "mcpServers": {     "reka": {       "command": "node",       "args": ["./mcp-server/dist/index.js"]     }   } }   # Index your codebase curl -X POST http://localhost:3100/api/index
Docker 24+ 8GB RAM Node 18+ ~2GB disk
OPEN SOURCE

Built in the open. Shaped by developers.

Reka is MIT licensed and always will be. No open-core bait-and-switch. Every feature, every line of code, available to everyone.

35+ MCP Tools
12k GitHub Stars
150+ Contributors
MIT License

Contribute

Good first issues tagged and waiting. PRs reviewed within 48 hours.

View open issues →

Discuss

Architecture decisions made in public. RFC process for major changes.

Join discussions →

Stay Updated

Monthly release notes. No spam, just shipped features.

Star on GitHub →