Option 1
It forgets you
Every session starts from nothing, so you explain your work again.
Personal AI OS Open source
Holt
An assistant that remembers, and shows you what it remembers.
Open-source personal AI OS and self-hosted AI assistant with memory
Holt is an open-source personal AI OS. Pick any LLM as the brain, keep your memory as files on your own machine, and open it as a graph whenever you want to check what it knows.
Holt is an open-source personal AI OS for people who work with an AI assistant: it keeps memory as files on your own machine, works with any LLM and lets you open that memory as a graph.
Price: Free, open source (MIT).
$ brew install holt-os/tap/holt # or, with Node already installed $ npm install -g @holt-os/holt
Free and open source, MIT licensed.
01 / The problem
Most assistants either forget you or keep a file on you that you cannot read.
Option 2
It profiles you in the dark
It remembers, but the profile sits on a server and you never get to read it.
Holt
Files you own
A graph you can walk and a wiki you can read and edit. Nothing about you is a black box.
02 / Persistent memory
Memory that lives in the folder you work in.
- →Per folder, kept between sessions. Launch Holt in a folder and the notes from earlier sessions there are already loaded.
- →It remembers on its own. One hook feeds the right past notes into each Claude Code prompt, then saves what mattered when the session ends.
- →Dated, so you can trust it. A recalled note comes back with the session it came from, so you can check it.
03 / Multi-brain
Pick the brain. Change it mid-conversation.
- →Use the CLI you already have. Bare holt launches Claude Code, Codex or Gemini, branded as Holt, with your memory wired in.
- →Or use an API key. Claude, OpenAI, Gemini, Kimi or Grok, with your own key.
- →The thread carries over. Switch brains and the same memory, skills and hooks run under the new one.
04 / Skills
Useful on the first day, not after a month of setup.
An empty assistant gives generic answers, and that is why most people install one and drift away. Holt ships thirteen skills, and three of them are a first-week loop.
- 01 /skill onboard
A fifteen minute interview that fills memory with your work, then does one real task.
- 02 /skill audit
Shows what Holt still does not know about you.
- 03 /skill level-up
Turns one manual chore a week into something Holt does for you.
- 13skills built in
- SKILL.mdportable agentskills.io format
- Registrysearch and add community skills
- MCPClaude Code, Cursor or Codex can read the memory
05 / Memory you can see
Open your memory and walk it.
- →A knowledge graph. holt graph opens an interactive map of what Holt knows and how it connects.
- →A cross-linked wiki. holt wiki keeps pages you can read, edit and link, and tends them for you.
- →Wrong note? Fix it. It is a file. Change it and the next session uses the change.
06 / Local and private
Your knowledge stays where you put it.
01
Memory per folder
What you tell Holt in one folder stays with that folder.
02
On your machine
Recall and storage run where you launched it.
03
Your own keys
You bring the API key or the CLI login. Holt does not sit in the middle.
04
Plain files
Open them, edit them, move them. No black-box profile.
05
Only the brain sees it
Your question and the notes it needs go to the model you picked, and nowhere else.
06
MIT licensed
Read the code, fork it, extend it.
07 / Memory you can check
Scored on the three public memory benchmarks.
Holt 0.19 on the public benchmarks for long-term AI memory. Every question, answer and grade is published, so you can rerun it.
91.2%
LongMemEval
500 questions about a user’s past chats. Graded by GPT-4o, the benchmark’s own grader. 89.7% on the 399 questions never used for tuning.
88.2%
LoCoMo
1,540 questions across long two-person conversations. Graded with the original, stricter LoCoMo prompt.
79.6%
BEAM, 1M tokens
700 questions over conversations of about a million tokens each.
Answers came from Claude Sonnet 5.5. LongMemEval was graded by GPT-4o, LoCoMo and BEAM by Claude Opus 5.5. All tuning used the same 101 LongMemEval questions. Full results and method →
08 / Install
Two minutes from install to a memory.
Install it, launch it in a folder, and run the onboarding skill. Not comfortable in a terminal? Paste the install guide into Claude Code and it sets Holt up for you.
# macOS, Linux, Windows $ brew install holt-os/tap/holt # or $ npm install -g @holt-os/holt $ cd ~/where-you-want-to-work $ holt # then, in that first session > /skill onboard
09 / For companies
Running a company? Look at Holt Teams.
Holt is for one person. Holt Teams is a separate, closed-source product that gives a whole team’s agents one shared memory, reviewed before it spreads, and kept inside the company. It also offers an offline option for regulated teams.
Questions
Questions about Holt
Is Holt free?
Yes. Holt is free and open source under the MIT license. You can install it with Homebrew or npm.
Does Holt work offline?
No. Holt runs on the LLM you pick, through a CLI such as Claude Code, Codex or Gemini, or with your own API key. Offline, local-model mode is part of Holt Teams only.
Which LLMs does Holt work with?
Bare holt launches Claude Code, Codex or Gemini with your memory wired in. Or use your own API key for Claude, OpenAI, Gemini, Kimi or Grok. You can switch brains mid-conversation and the memory carries over.
Where does Holt store my memory?
As files on your own machine, kept per folder between sessions. You can open it as a knowledge graph with holt graph, or as a wiki you can read and edit.
How do I install Holt?
Run brew install holt-os/tap/holt, or npm install -g @holt-os/holt, then launch holt in a folder and run the onboarding skill.
What is the difference between Holt and Holt Teams?
Holt is for one person and is open source. Holt Teams is a separate, closed-source product that gives a whole team's agents one shared memory, kept inside the company.
Give your assistant a memory you can read.
Free, open source, and running on your machine in a couple of minutes.
← Back to all products