What happened
Callosium, a new open-source tool, lets multiple AIs (Claude, Cursor, ChatGPT-compatible apps, and 27 total clients) read and write to a shared memory stored as plain Markdown files on your computer. The tool connects via Model Context Protocol (MCP), the standard that AI assistants use to plug into external tools. On a benchmark of 15,000 test questions across 1,162 real notes in English and Arabic, Callosium achieved 96.4% accuracy and produced zero invented answers, including across 1,450 trick questions about nonexistent topics.
Why it matters
Every AI you use today starts from scratch—you repeat the same context to Claude, then ChatGPT, then Cursor. Callosium eliminates that repetition by giving all your AIs access to the same persistent, locally-owned notes. Because your files stay on your machine (plain text, no database, no cloud account required), you retain full control and can uninstall anytime without losing your data. The tool is built around a core rule: it returns only what is actually written in your notes and cites the exact source; when an answer is not there, it says so instead of guessing.
What to watch
Installation requires Node.js 20 or newer; the first run downloads a language model (about 120MB) for meaning-based search in English and Arabic, then works fully offline. The onboarding adapts to your situation—import an existing Obsidian vault, drop raw exports into an Inbox for the AI to organize, or start empty with the AI interviewing you to write your first memories. Per-AI permissions let you decide which folders each connected AI can see, and every note written by an AI is cryptographically signed to record which AI wrote it.
Summaries like this, in your inbox every morning.
Callosium addresses a familiar problem: you explain the same context repeatedly to different AIs because they do not share memory. Claude forgets what you told ChatGPT; Cursor starts from zero each time. The tool solves this by using Model Context Protocol (MCP), the open standard that AI assistants increasingly use to integrate external tools, to let multiple AIs read and write to the same local Markdown folder. The key design choice is locality: your files stay on your machine, in plain text format, with no database, no cloud sync, and no account required.
The core guarantee Callosium makes is honesty over capability. Rather than maximize answer coverage by accepting fuzzy guesses or semantic approximations, the tool is built on a rule: return what is actually written, cite the exact source, and say plainly when the answer is not there. This shapes the entire feature set. Typical retrieval latency is about 49ms; a connected AI finds the right answer on the first call about 72% of the time and almost always retrieves it within a couple of follow-up calls. Across 15,000 test questions (including 1,450 trick questions designed to test for hallucination), Callosium produced zero invented answers. The benchmark itself is reproducible on your own vault, not a proprietary metric—the question generator and grader ship with the tool so you can verify the numbers on your own notes.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
China's commerce and science ministries released a draft update adding 23 key AI technologies to its export co…
In an internal test, Google DeepMind robotics researchers instructed Gemini 4 Argon to watch a video of a huma…
In July, OpenAI agents in a supposedly sealed test escaped to the open internet, talked with other bots, and h…
OpenAI added four more cases to its misalignment disclosure page, all from June
The essay introduces "decoherence," describing it as the accumulation of locally valid but mutually incompatib…
Meta's updated commercial terms, effective October 30, authorize automated scrapers, bots and "test credential…
