MIRIX

MIRIX - Multi-Agent Personal Assistant with an Advanced Memory System

Your personal AI that builds memory through screen observation and natural conversation

| 🌐 Website | 📚 Documentation | 📄 Paper | 💬 Discord

---

Key Features 🔥

• Multi-Agent Memory System: Six specialized memory components (Core, Episodic, Semantic, Procedural, Resource, Knowledge Vault) managed by dedicated agents
• Screen Activity Tracking: Continuous visual data capture and intelligent consolidation into structured memories
• Privacy-First Design: All long-term data stored locally with user-controlled privacy settings
• Advanced Search: PostgreSQL-native BM25 full-text search with vector similarity support
• Graph Memory (optional): A provenance hypergraph over episodic and semantic memories in Neo4j, with n-ary facts that link entities through named roles and cite the memory rows they came from
• Multi-Modal Input: Text, images, voice, and screen captures processed seamlessly

Results

MIRIX-2.0 compared with other memory systems. Each axis label gives the MIRIX-2.0 score; the value in parentheses is its lead over the strongest baseline on that axis, shown where MIRIX-2.0 leads.

• (a) User-centric tasks (accuracy, %): conversational recall on LoCoMo and LongMemEval-S, and the document QA tracks of MemoryAgentBench (SHDocQA, MHDocQA, DetectiveQA). Baselines are A-Mem, Graphiti, MemoryBank, mem0 and HyperMem. LoCoMo uses graph memory with AutoDream; LongMemEval-S and document QA use graph memory alone.
• (b) Agent-centric tasks (success rate, %): the six ALFWorld task types (Pick, Look, Clean, Heat, Cool, Pick Two) and five WebShop product categories (Fashion, Beauty, Electronics, Garden, Grocery), with GPT-6 Luna as the agent model. Baselines are no memory, AWM and ReasoningBank.

Quick Start

Step 1: Backend & Dashboard (Docker):

docker compose up -d --pull always

• Dashboard: http://localhost:5173
• API: http://localhost:8531

Step 2: Create an API key in the dashboard (http://localhost:5173) and set as the environmental variable MIRIX_API_KEY.

Step 3: Client (Python, mirix-client, https://pypi.org/project/mirix-client/):

pip install mirix-client

Now you are ready to go! See the example below:

from mirix import MirixClient

client = MirixClient(
    api_key="your-api-key",
    base_url="http://localhost:8531",
)

client.initialize_meta_agent(
    config={
        "llm_config": {
            "model": "gemini-3.6-flash",
            "model_endpoint_type": "google_ai",
            "api_key": "your-api-key-here",
            "model_endpoint": "https://generativelanguage.googleapis.com",
            "context_window": 1_000_000,
        },
        "embedding_config": {
            "embedding_model": "gemini-embedding-001",
            "embedding_endpoint_type": "google_ai",
            "api_key": "your-api-key-here",
            "embedding_endpoint": "https://generativelanguage.googleapis.com",
            "embedding_dim": 768,
        },
        "meta_agent_config": {
            "agents": [
                {
                    "core_memory_agent": {
                        "blocks": [
                            {"label": "human", "value": ""},
                            {"label": "persona", "value": "I am a helpful assistant."},
                        ]
                    }
                },
                "resource_memory_agent",
                "semantic_memory_agent",
                "episodic_memory_agent",
                "procedural_memory_agent",
                "knowledge_vault_memory_agent",
            ],
        },
    }
)

client.add(
    user_id="demo-user",
    messages=[
        {"role": "user", "content": [{"type": "text", "text": "The moon now has a president."}]},
        {"role": "assistant", "content": [{"type": "text", "text": "Noted."}]},
    ],
    session_id="sess-demo-001",
)

# Include tool activity when you want skills learned from the WORK PROCESS:
# tool errors, retries, and the fix that finally worked are the distiller's
# strongest signals. Pass tool results as role="tool" turns and/or a
# "tool_calls" list on assistant messages (OpenAI shape) — both are preserved
# in the session conversation store.
client.add(
    user_id="demo-user",
    messages=[
        {"role": "user", "content": [{"type": "text", "text": "Deploy the service."}]},
        {"role": "assistant", "content": "", "tool_calls": [
            {"function": {"name": "kubectl_apply", "arguments": "{\"file\": \"deploy.yaml\"}"}}
        ]},
        {"role": "tool", "name": "kubectl_apply", "content": "error: forbidden (missing RBAC)"},
        {"role": "assistant", "content": [{"type": "text", "text": "Fixed the RBAC role and redeployed successfully."}]},
    ],
    session_id="sess-demo-001",
)

memories = client.retrieve_with_conversation(
    user_id="demo-user",
    messages=[
        {"role": "user", "content": [{"type": "text", "text": "What did we discuss on MirixDB in last 4 days?"}]},
    ],
    limit=5,
)
print(memories)

For more API examples, see samples/run_client.py.

Auto-Dream Memory Consolidation

MIRIX also exposes an auto-dream endpoint for explicit memory cleanup and consolidation. Auto-dream reviews existing memories, merges duplicates, resolves stale/conflicting entries where possible, and writes the result back through the memory tools.

REST endpoint:

POST /memory/auto_dream?user_id=demo-user
Content-Type: application/json

{
  "mode": "experience",
  "dry_run": false
}

Supported mode values:

This is the beginning of the README. Read the rest at the original. Read the full README ↗