Chat
Claude Code in an interface that keeps up: tool calls in the transcript, attachments, model choice, several conversations per project.
An assistant that remembers.
Osiris gives Claude Code a memory: a knowledge graph in Markdown that holds what applies in your projects — and a long-term memory that learns from every conversation instead of forgetting it.
Get started
How to install
macOS, Linux & Windows · needs Node 22 · opens in its own window
curl -fsSL https://osirisui.com/install | sh
Then start it anytime with osiris start.
Or run from source (for developers)
# Requires Node 22 and the Claude Code CLI $ git clone https://github.com/TheRealCequu/OSIRIS.git $ cd OSIRIS/software $ node bin/osiris.mjs start Interface http://127.0.0.1:7777
A language model forgets everything after every chat. That is the actual defect.
Not the quality of the answers — those are fine. It is that you explain the same rules, the same decisions, the same quirks of the project from scratch every single time. Osiris fixes exactly that, in two places: once for what applies, and once for what happened.
Ultrabrain
Write it down once, and it holds everywhere
“Never store money as a float.” “Menus are chest inventories, back
button bottom left.” Sentences like these you otherwise explain to a model in every
second chat. In the Ultrabrain they live as nodes: Markdown with a header, linked to
one another, assigned to a project.
- Claude reads them before the first decision that matters — without you mentioning them.
- One click writes the applicable rules into the repo's
CLAUDE.md. - Markdown stays Markdown: in Obsidian, in Git, and without Osiris.
Ultrabrain
GraphListAll projectsBladeStudios
Graph
129Nodes
214Edges
2Projects
Memory
Compiled on write, not on every question
Ordinary RAG systems cut documents into chunks and search by similarity on every question. Osiris does it the other way round — following the pattern Andrej Karpathy described in April 2026 as the “LLM wiki”: after every conversation a separate run reads it and condenses it into pages. What gets searched is then not raw material, but something that has already been understood.
- Ingest — ten minutes after the last turn, once per conversation.
- Query — the map first, then precisely the pages that fit.
- Maintain — weekly: contradictions, orphans, gaps.
Memory
MapPagesSources
Memory
36Pages
10Sources
0open
raw/What happenedEvery conversation, every file put there — unchanged. The agent reads them, it never rewrites them.
wiki/What came of itEntities, concepts, comparisons, answers worth keeping. Every paragraph cites its source.
schema.mdHow it is writtenThe rules of the memory — as a file you can read and change.
log.mdWhat was doneOne line per operation, append-only. Nothing happens unnoticed.
Measured, not believed
A memory needs a benchmark
25 pages, 30 questions, each with a known answer page. Every change to retrieval therefore has a number before and a number after. These come from exactly that run.
90 → 97 %Recall@5 — the right page among the first five
0.86 → 0.93MRR — how far up it ranks
33 → 100 %Compound words: “Tickbudget” finds “Tick-Budget”
0Dependencies — Node built-ins, no vector store
Abilities
Forms instead of prompts
Some tasks you do over and over — and phrase a little differently every time. An ability is such a task as a form: fill in the fields, submit. The instruction that reaches Claude is the same every time, with the same rules from the Ultrabrain behind it.
Minecraft Development
Create and open a plugin project
Create and open a plugin project
Paper 1.21
Vault
Adventure
Gradle
What Claude makes of it
created BladeShop/build.gradle.kts created src/main/resources/plugin.yml created …/shop/BladeShop.java created …/shop/menu/ShopMenu.java created …/shop/wirtschaft/VaultBruecke.java Applied from the Ultrabrain: · Money in cents, never as a double · Menu 9×6, back bottom left, cancel the click · No file access on the main thread · Messages with prefix, colours with meaning Project opened in the IDE.
The example is real: it is the ability that creates the Minecraft plugins at BladeStudios. Abilities are text files — writing your own takes a few lines.
And also
What else is in the box
Explorer
The whole knowledge tree to browse and edit — with preview, links and backlinks, without switching to the editor.
Projects
Link repos, preview the context beforehand, and write the applicable rules into CLAUDE.md — where Claude Code reads them by itself.
Command line
Windowless, too
Everything the interface can do, the terminal can do as well: start, stop, search, look up, sync, check the state. And the wordmark carries the colour you set in the app.
$ osiris doctor ██████ ██████ ██████ ██████ ██████ ██████ ██▒▒██▒██▒▒▒▒▒ ▒██▒▒▒██▒▒██▒ ▒██▒▒▒██▒▒▒▒▒ ██▒ ██▒██████ ██▒ ██████▒ ██▒ ██████ ██▒ ██▒ ▒▒▒██▒ ██▒ ████▒█▒ ██▒ ▒▒▒██▒ ██▒ ██▒ ██▒ ██▒ ██▒▒██▒ ██▒ ██▒ ██████▒██████▒██████ ██▒ ██▒██████ ██████▒ ▒▒▒▒▒▒ ▒▒▒▒▒▒ ▒▒▒▒▒▒ ▒▒ ▒▒ ▒▒▒▒▒▒ ▒▒▒▒▒▒ Ultrabrain for Claude Code ● Internet Name resolution succeeded ● Anthropic API Reachable (130 ms) ● Claude Code CLI 2.1.278 ● Ultrabrain 129 nodes, 214 links ● Memory 36 pages, 10 sources, 0 open Overall: ● ok
Where your data lives
On your disk. Full stop.
Osiris is a program that runs on your machine — no cloud, no account with us, no server reading along. Your knowledge graph and your memory are folders full of Markdown files that belong to you.
No account with us
Osiris talks to Claude through Anthropic's Claude Code CLI — with your own access. We are not in between and see none of it.
What goes to Anthropic
Only what you write in the chat, together with the context Osiris adds. The preview shows you every character of it beforehand.
This website
No Google fonts, no CDN, no analytics, no embeds. Two settings are stored, and only when you click them: light or dark, and your language.