A modular multi-agent memory system that can be plugged into any LLM agent stack. Specialized agents manage six memory types (Core, Episodic, Semantic, Procedural, Resource, Knowledge Vault) under a meta-agent coordinator that orchestrates writes and retrieval. Backed by PostgreSQL with BM25 full-text search and embedding-based vector similarity. Exposed as a self-hosted REST API or via managed hosting at app.mirix.io. The screen-activity tracking desktop app was deprecated in v0.1.6 and lives on a separate branch.
- Storage
- PostgreSQL-backed stores for the six memory categories, combining BM25 full-text search and vector similarity indexing; deployable via Docker Compose or local install.
- Retrieval
- A meta-agent coordinator routes incoming content to the relevant memory-type agents, orchestrates writes, then surfaces results through a unified REST API or Python SDK; query by conversation context, keyword, or embedding similarity.
- Self-host
- Self-host: moderate
- License
- Apache-2.0
- Pricing
- Open source (Apache-2.0), free to self-host via Docker or local install; managed hosting available at app.mirix.io (pricing not publicly listed). · Freemium
- GitHub stars
- 3,547
- Last release
- 2025-12-25
- Last commit
- 2026-06-20
- First catalogued
- 2026-06-28
Strengths
- Multi-agent design with six distinct memory types covering the full range of personal and semantic memory
- Model-agnostic — works with OpenAI, Google AI, or any OpenAI-compatible endpoint
- PostgreSQL-backed storage with BM25 full-text search plus embedding-based vector retrieval
- Pluggable into any agent stack via REST API or Python SDK
- Auto-dream memory consolidation endpoint for deduplication and staleness resolution
Watch out
- Desktop personal-assistant and screen-tracking app deprecated in v0.1.6; now on a separate branch — card describes the main-branch API
- GitHub releases stop at v0.1.6; newer PyPI versions (0.1.7+) may not have formal release notes
- Managed hosting at app.mirix.io has no publicly listed pricing
- Self-reported LoCoMo/ScreenshotVQA numbers exist — route to harvest-benchmarks, do not treat as neutral here
Best for
- Agent developers needing a drop-in, multi-type memory layer behind a REST API
- Projects requiring BM25 + vector hybrid retrieval across distinct memory categories
- Teams that want local-first, self-hosted storage with an optional managed fallback
How it integrates
Benchmark results
| Benchmark | Value | Backbone | Trust | Source |
|---|---|---|---|---|
| locomo | 85.38 accuracy | gpt-4.1-mini | Self-reported | MIRIX (Wang & Chen) ↗ |
| longmemeval | 43.49 accuracy | GPT-4o-mini | Independent | MemOS (MemTensor) — competitor re-run ↗ |
Sources
- Monthly card re-verification — GitHub API repo metadata: 3,558 stars, license Apache-2.0 confirmed, archived=false, last commit on the default branch 2026-06-20 (third-party)
- https://github.com/Mirix-AI/MIRIX (vendor)
- MIRIX: Multi-Agent Memory System for LLM-Based Agents (paper)
- MIRIX Documentation — Overview (vendor)
- MIRIX Documentation — Quickstart (vendor)
- mirix on PyPI (v0.1.7) (vendor)
- MIRIX GitHub Releases (latest: v0.1.6, 2025-12-25) (vendor)
- GitHub license API — confirms Apache-2.0; repo not archived, pushed 2026-07-15 (third-party)
- MIRIX README (current main) — Docker Compose quickstart, `pip install mirix-client` command (package not actually published; real PyPI name is 'mirix'), auto-dream endpoint unchanged (vendor)
Last verified 2026-08-01 · updated by refresh-framework-cards