An AI that remembers what you did on screen last Tuesday

What it does

Mirix is a self-hosted personal assistant that captures screen activity and multimodal inputs—text, images, voice, and screen grabs—and consolidates them into a structured, queryable memory store. It exposes a local dashboard and API, letting you ask natural-language questions about your past digital activity without sending your history to a cloud vault.

The interesting bit

Rather than treating memory as a single vector dump, Mirix divides it into six agent-managed buckets—Core, Episodic, Semantic, Procedural, Resource, and Knowledge Vault—built on the open Letta framework. The bet is that a taxonomy of memory types makes retrieval more precise than one giant embedding search.

Key highlights

• Six specialized memory agents manage distinct memory types, from core identity to episodic events and procedural knowledge.
• All long-term data is stored locally with user-controlled privacy settings; the stack runs as a containerized backend.
• Search combines PostgreSQL-native BM25 full-text with vector similarity.
• Accepts text, images, voice, and screen captures as inputs.
• The memory system architecture derives from the Letta framework, which the authors acknowledge as their foundation.

Caveats

• The sample configuration routes LLM inference to external APIs (Google AI), so while storage is local, the system is not fully offline by default.
• “Continuous visual data capture” is claimed, but the README offers no detail on capture frequency, OCR, or on-device processing.
• How the six memory agents coordinate and avoid duplication is not explained in the provided sources.

Verdict

Worth a look if you want a local memory server for LLM assistants and are comfortable wiring up your own model endpoints. Skip it if you need a turnkey offline LLM stack or deep transparency into the screen-tracking pipeline.