MIRIX
Multi-agent personal assistant framework that builds structured long-term memory from screen activity and natural chat.
Track screen activity, build provenance graph memory, and consolidate agent context with Auto-Dream.
MIRIX is an open-source personal AI assistant framework that turns continuous screen observation and chat logs into long-term structured memory. It uses six specialized memory agents to categorize facts, procedural steps, and resources. You should use it if you want to build personal productivity agents that remember past conversations, workflows, and screen context.
What it is
MIRIX captures visual data, voice, and text, converting raw activity into six memory types including core, episodic, and procedural stores. It indexes memories with PostgreSQL BM25 full-text search, vector embeddings, and an optional Neo4j provenance hypergraph. An Auto-Dream consolidation engine actively deduplicates memories and extracts learned skills from multi-step tool interactions.
Who it's for
This tool suits developers and systems engineers building context-aware executive AI assistants that require deep long-term recall across sessions. It works best for teams comfortable managing Docker containers, Python client SDKs, and external LLM endpoints like Gemini. It does not suit non-technical users looking for a desktop app with zero setup.
Key features
Six Specialized Memory Agents
Organizes captured context into Core, Episodic, Semantic, Procedural, Resource, and Knowledge Vault stores.
Auto-Dream Memory Consolidation
Reviews existing entries to merge duplicates, resolve stale facts, and distill procedural skills from tool logs.
Provenance Hypergraph in Neo4j
Indexes episodic and semantic facts as n-ary frames in a graph, citing the precise memory rows they originated from.
PostgreSQL Vector and BM25
Combines native BM25 full-text search with vector embeddings for structured recall queries.
Up and running in 3 steps
Our honest verdict
What's great
- Distills procedural rules directly from tool call errors and retries in conversation sessions.
- Supports dry-run modes for memory consolidation before applying updates to persistent stores.
- Stores long-term data locally with user-controlled privacy settings.
- Provides optional graph memory backends that preserve exact provenance for extracted facts.
Where it falls short
- Requires running Docker containers, PostgreSQL, and an optional Neo4j 5.23+ cluster.
- Demands active API keys for external LLMs and embedding providers such as Google AI Gemini.
- Requires technical configuration using Python SDK calls rather than offering a simple GUI installer.
Pricing & license
MIRIX is completely free under the Apache-2.0 open-source licence. The project documentation mentions no paid cloud subscription, meaning your only costs come from hosting the Docker infrastructure and paying for third-party LLM API calls.
Open remote AI-training roles
873 open remote AI-training roles, updated November 15, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
What is the Auto-Dream feature in MIRIX?
Auto-Dream is an explicit memory maintenance endpoint that reviews past memories, resolves conflicting entries, merges duplicates, and distills procedural skills from session logs.
Does MIRIX require Neo4j to function?
No, Neo4j is optional. MIRIX works out of the box with PostgreSQL vector and BM25 search, but you can enable Neo4j if you need provenance hypergraphs.
How does MIRIX learn procedural knowledge?
When you record conversation turns with tool execution errors and successful fixes, MIRIX distills those work process attempts into procedural memory.
