# memxt

> Local-first long-term memory for AI coding agents

The fastest benchmarked open-source AI memory system — a single static binary that gives coding agents persistent, on-device memory across sessions. No cloud, no Python, nothing leaves your machine.

## Built for

Developers who want coding agents to remember architecture, decisions, and working context across sessions without sending that memory to a cloud service.

## Product thesis

Your Agent Can Forget the Chat. It Should Not Forget the Work.

## Capabilities

- **Wake Up with Context:** MCP tools and session hooks give an agent relevant project memory at the start of a session instead of another blank slate.
- **Keep Recall Out of the Way:** A Zig binary, local embedding model, llama.cpp, and sqlite-vec keep semantic search fast enough to become invisible infrastructure.
- **Keep Decisions On-Device:** Embeddings, storage, and search stay on the machine. There is no hosted vector database and no API key at query time.

## Evidence

- ~10 ms session wake-up
- <1 ms warm vector search
- 1 binary no Python or Docker

## Details

- Status: live
- Category: DeveloperApplication
- Operating systems: macOS, Linux, Windows
- Stack: Zig, LLM, MCP, AI Memory
- Keywords: memxt, memxt agent memory, AI agent memory, long-term memory for AI agents, MCP memory server, local-first AI memory, Palash Debnath
- Open source: yes
- Updated: 2026-08-13
- Pricing: free
- GitHub stars: 33
- Release downloads: 67
- Latest release: [v0.4.0](https://github.com/Yupcha/memXT/releases/tag/v0.4.0) — 2026-07-18
- License: MIT
- Website: https://yupcha.github.io/memXT/
- Source: https://github.com/Yupcha/memXT

Canonical page: https://palash.dev/memxt/