MIRIX Redefines Multimodal Long-Term AI Memory: 410% Super Gemini, 99.9% Memory Savings, APP Launch}

MIRIX, a new system led by UCSD and NYU teams, is redefining AI memory architecture.

Over the past decade, large language models have swept the globe, transforming from writing assistants to code generators. Yet, even the most powerful models have a fundamental weakness: they do not remember you.

Addressing this, UCSD PhD student Yu Wang and NYU professor Chen Xi (Xi Chen) jointly launched and open-sourced MIRIX — the world's first truly multimodal, multi-agent AI memory system.

• Paper Title: MIRIX: Multi-Agent Memory System for LLM-Based Agents
• Official Website: https://mirix.io/
• Open Source Repository: https://github.com/Mirix-AI/MIRIX

MIRIX's performance is remarkable! On the ScreenshotVQA benchmark, which requires deep multimodal understanding, MIRIX outperforms traditional RAG methods by 35%, with 99.9% less storage. Compared to long-text methods, it exceeds by 410% and reduces storage by 93.3%. In LOCOMO multi-turn dialogue reasoning, MIRIX achieves 85.4%, setting new performance standards.

Additionally, a Mac desktop app has been launched, allowing anyone to see, understand, and convert their experiences into persistent digital memories.

A Completely New Paradigm

Reflecting on AI evolution over the past three years:

• Limits of large model reasoning: parameter explosion & compute surge;
• Vector retrieval as emergency: RAG fragment stitching;
• Introduction of "short-term memory": limited dialogue history.

But none of these are true "mind systems." MIRIX first achieves:

1. Multimodal input support: understanding not just text but also high-res screenshots, logs, and multi-source data to build global memory.
2. Six types of human memory systems: each with dedicated data structures, lifecycle management, and retrieval routing.
3. Built-in multi-agent collaboration: via a Meta Memory Manager overseeing six parallel memory agents, working together with a unified dialogue agent.
4. Proactive topic generation & layered retrieval: analyzing user intent first, then generating topic embeddings, matching memory types for multi-layer retrieval.
5. Product deployment: a personal assistant app now available on Mac, enabling instant multimodal data collection and visualized hierarchical memory management, stored locally in SQLite for privacy.

It no longer just embeds knowledge into "vector spaces" but constructs a structured, multi-channel, evolvable cognitive base.

Why is it considered an AI "proto-mind"?

MIRIX features six core memory types, enabling nuanced cognitive roles:

• Core Memory: Stores AI's "personality" and user preferences—dialogue style, settings, identity—persistently as KV pairs, auto-loaded in responses.
• Episodic Memory: Like human event logs, with timestamps, event types, subjects, summaries, and details, retracing all user interactions.
• Semantic Memory: Stores concepts, facts, and social graphs, with entries as tuples of name, definition, description, and source, supporting multi-hop reasoning.
• Procedural Memory: Saves task workflows as JSON-structured multi-step operations, e.g., "How to fill expense reports".
• Resource Memory: Stores complete files or snippets, supporting cross-task referencing, e.g., uploaded contracts or meeting notes.
• Knowledge Vault: Sensitive info like passwords, API keys, encrypted with multi-level access control.

Unlike RAG, which fetches all info at once, MIRIX understands what to recall, then searches and combines answers—thinking about "what to remember" rather than just indexing.

Multi-Agent Workflow

Handling long-term, dynamic, multimodal data requires a modular multi-agent architecture, with dedicated components collaborating under a unified scheduler: Meta Memory Manager, Memory Managers, and Chat Agents.

When new input arrives (text, inferred events, files), the system:

1. Preliminary retrieval: searches existing memory for similar records to avoid redundancy and update existing entries.
2. Routing & analysis: parses retrieved info and raw input, extracting metadata and deciding which memory modules to update.
3. Parallel update: each memory manager extracts structured info, deduplicates, and updates indexes and embeddings.
4. Confirmation: once updates complete, the system confirms the memory update is finished.

Performance Breakthrough — Crushing RAG and Other Memory Systems

In the new multimodal challenge ScreenshotVQA, MIRIX reduces storage by 99.9% while achieving 35% higher accuracy than RAG; surpasses full-context reasoning with 410% improvement and 93.3% less storage. In LOCOMO multi-turn reasoning, MIRIX scores 85.4%, far exceeding baselines like MemOS and Mem0, nearly reaching GPT-4.1-mini's upper limit with a 1M context window.

This is not just "better"—it’s a completely different paradigm

As MIRIX team states: "Retrieval-augmented generation is just a patch. True memory allows AI to grow over time." This marks a new cycle: from "dialogue generation" to "long-term memory-driven cognition".

Moreover, MIRIX is not just a paper; the team has launched a desktop personal assistant app that enables real-time multimodal data collection, visual hierarchical memory management, and local storage in SQLite to protect privacy. Download now to experience AI that truly "remembers you".