openeruka

The open-source knowledge state memory server for AI agents.

openeruka is a fully self-contained memory server that enforces the knowledge state invariant: Confirmed facts cannot be overwritten by Inferred guesses. Run it locally, connect eruka-mcp to it, and your AI agent gets grounded, protected memory. Ships as a single binary with SQLite bundled — production-ready in one command.

# Install the server binary
cargo install openeruka

# Start it
openeruka serve                  # default port 8080

# Connect Claude Code / Claude Desktop via eruka-mcp
# Default ERUKA_API_URL is http://localhost:8080 — no extra config needed

The knowledge state invariant

use openeruka::KnowledgeState;

// The core contract — enforced server-side, not advisory
assert!( KnowledgeState::Confirmed.can_overwrite(&KnowledgeState::Inferred));
assert!(!KnowledgeState::Inferred.can_overwrite(&KnowledgeState::Confirmed));

When your agent tries to overwrite a Confirmed fact with an Inferred guess, the server returns 409 Conflict. LLM hallucinations cannot corrupt verified knowledge. No other memory system enforces this.

Interaction surfaces

openeruka exposes three equivalent interfaces:

Interface	Command	Use case
REST API	openeruka serve	Connect any HTTP client, agent, or openeruka-client
MCP server	openeruka mcp	Claude Desktop, Claude Code, Cursor via eruka-mcp
CLI	openeruka get/set	Quick reads and writes from the terminal

All three share the same SQLite backend and enforce the same knowledge state rules.

Quick start

# Install
cargo install openeruka

# Start REST server (SQLite backend, default ./eruka.db)
openeruka serve

# Write a confirmed fact
curl -X POST http://localhost:8080/api/v1/context \
  -H "Content-Type: application/json" \
  -d '{"workspace_id":"my-project","path":"identity/company_name","value":"Acme Corp","knowledge_state":"CONFIRMED","confidence":1.0,"source":"user_input"}'

# Try to overwrite with an inferred guess — rejected with 409
curl -X POST http://localhost:8080/api/v1/context \
  -H "Content-Type: application/json" \
  -d '{"workspace_id":"my-project","path":"identity/company_name","value":"Acme AI Labs","knowledge_state":"INFERRED","confidence":0.7,"source":"agent_inference"}'
# → 409 Conflict: field is CONFIRMED, write has lower knowledge state

# Read it back — still the original
curl "http://localhost:8080/api/v1/context?workspace_id=my-project&path=identity/company_name"

Use as a library

[dependencies]
openeruka = { version = "0.2", default-features = false }
openeruka-client = "0.1"

// Embed SqliteContextStore in your own Rust agent
use openeruka::{SqliteContextStore, ContextStore, ErukaFieldWrite, KnowledgeState, SourceType};

let store = SqliteContextStore::open("./eruka.db").await?;

// Write enforces the invariant
store.write_field("my-ws", ErukaFieldWrite {
    path: "identity/name".to_string(),
    value: serde_json::json!("DIRMACS"),
    knowledge_state: KnowledgeState::Confirmed,
    confidence: 1.0,
    source: SourceType::UserInput,
}).await?;

openeruka vs eruka.dirmacs.com

eruka-mcp can connect to either:

# Local openeruka (OSS, SQLite, single-tenant)
ERUKA_API_URL=http://localhost:8080 eruka-mcp

# Managed Eruka (PostgreSQL, multi-tenant, enterprise features)
ERUKA_API_URL=https://eruka.dirmacs.com ERUKA_API_KEY=sk:your-key eruka-mcp

	openeruka (OSS)	eruka.dirmacs.com
Knowledge state enforcement	✅	✅
REST + MCP + CLI	✅	✅
SQLite (local)	✅	—
PostgreSQL (scalable)	—	✅
Multi-tenancy	—	✅
B6 quality scoring	—	✅
Datalog inference	—	✅
Gardener pipeline	—	✅
SLA / HIPAA / SSO	—	✅

DIRMACS ecosystem

• eruka-mcp — MCP client for openeruka and eruka.dirmacs.com
• ARES — multi-agent runtime that uses Eruka for context
• pawan — CLI coding agent with Eruka memory
• deagle — code intelligence engine
• DIRMACS — the company behind this stack

Docs

dirmacs.github.io/openeruka

License

MIT — see LICENSE