What Is Recursive Human Intelligence? The Beginner's Guide to LLM Chat Analytics

CodexFuse Team · 2026-11-05T00:46:22.167Z · 11 min read min read

What Is Recursive Human Intelligence?

Recursive Human Intelligence (RHI) is a free, open-source desktop app that reads your Claude Code, Codex, and Pi chat logs and gives you blunt, actionable feedback on your own LLM habits — what you keep getting stuck on, and what would fix it for good. Built by developer Andrea Ritossa and launched on November 3, 2026, it starts from a simple premise, quoted verbatim from its README: "AI improves itself recursively. Why shouldn't you?"

The idea is to treat your chat history as training data for yourself. Instead of summarizing individual conversations, RHI groups your requests into recurring behavior patterns and shows you, with your own words as receipts, where you waste time: the prompt you retype in 23 different chats, the complex model you reach for on simple tasks, the decision you keep reopening. It ships as a desktop app for your own machine — the same kind of always-on box a VPS hosting plan gives your side projects — and it works only with chats you already made. It analyzes; it does not run agents or chat on your behalf.

What Does RHI Actually Read?

RHI reads local chat logs from three platforms: Claude Code, Codex, and Pi. Those are the only sources it understands today, and it reads the logs already sitting on your disk — you do not paste conversations into it or connect any accounts.

What comes out is not a usage dashboard. The README puts it this way: "Not a usage dashboard. Feedback a blunt friend would give after reading every request you ever typed." Every item it surfaces quotes your own messages, so you can check its work instead of taking its word for it. For Claude chat analytics drawn from Claude Code sessions specifically, that means RHI can show you patterns across months of pairing sessions — like over-explaining context in every session when a system prompt would have sufficed.

How Does Recursive Human Intelligence Work?

The whole loop is five steps, and the interface is famously minimal — the README says "Click Get feedback. That's the whole interface." Here is the arc from zero to first insight:

1. Install it. One command for your OS (see the install section below). No Java required; the installer bundles its own runtime.
2. Connect a model. Paste an API key for OpenAI, Anthropic Claude, or OpenRouter in Settings, or set the matching environment variable, then click Test connection and allow excerpt sending.
3. Click Get feedback. One click starts a complete read: the analyzer keeps looking, up to ten passes, grouping your requests into recurring behaviors live until nothing strong is left.
4. Act on what arrives. Open a feedback item, do the one thing it asks — save the templated prompt, settle the decision, learn the concept — and mark it done. New chats unlock new feedback.
5. Let the Sunday pass run. Every Sunday at 16:00 on your computer's clock, while RHI is running, it reads new chats, generates fresh feedback, and sends a desktop notification.

History and state live in a local ~/.loop/ folder — history.tsv, scan-index.tsv, settings.properties, solutions.json, and generated HTML artifacts. Your transcripts are never changed by the app.

What Kind of Feedback Does It Give You?

RHI delivers three kinds of feedback, and all three come with quoted receipts from your own messages. First, your habits, with receipts: the README's own example is "You've typed this phrase in 23 chats. Save it once, with real limits." It shows what you wrote and a sharper version.

Second, what to actually learn. If you keep asking about the same subject, it names a real book or course instead of answering again — the point being that some questions are cheaper to learn once than to re-ask forever. Third, what you keep reopening: decisions you never settle, and questions you are asking an AI that only your customers can answer.

Launch coverage adds four pattern types the tool hunts for: repetitive prompt structures that could be automated, escalating to complex models for simple tasks, opportunities to reuse proven interaction patterns, and how your conversation style evolves over time. Together they form the LLM chat analytics core of the app — less "how much did you chat" and more "how could you chat better."

How Do You Install Recursive Human Intelligence?

Installing the Recursive Human Intelligence app takes one command, copied from the Recursive Human Intelligence GitHub repo. The recursive human intelligence install is a single terminal command per OS. On macOS or Linux, run curl -fsSL https://raw.githubusercontent.com/andrearitossa/recursive-human-intelligence/main/install.sh | sh in a terminal. On Windows, open PowerShell and run irm https://raw.githubusercontent.com/andrearitossa/recursive-human-intelligence/main/install.ps1 | iex.

Two things to know before you run it. Builds are not code signed yet, so your system may warn you — the README advises reading install.sh or install.ps1 first, noting the scripts are short and install only for your user. And if you would rather skip scripts, the repo's Releases page carries prebuilt binaries for every platform (four releases exist as of November 5, 2026).

Building from source is the only path that needs Java: JDK 17 or newer, then desktop/start.sh (or start.bat on Windows) compiles, runs checks, and opens the app. A desktop/check.sh validation script is included for contributors.

What Does "Bring Your Own LLM" Mean?

RHI ships with no built-in model and no free tier — analysis calls go to an LLM provider you choose and pay for. Settings supports three providers through one model interface: OpenAI (OPENAI_API_KEY), Claude via Anthropic (ANTHROPIC_API_KEY), and OpenRouter (OPENROUTER_API_KEY).

You paste a key or set the environment variable, enter a model ID your account can access, and click Test connection. Keys are never written to settings — they are encrypted by the Windows credential store, macOS Keychain, or the Linux keyring, or kept in memory for the session if no keyring exists. A "Forget saved key" button in Settings removes a stored key. Custom endpoints are allowed as long as they use HTTPS, with an exception for localhost development endpoints.

This is the practical meaning of the launch coverage's "bring your own LLM" line: you control which provider sees your excerpts, and you can switch providers without reinstalling. No key means no analysis — the app reads your chats locally for free, but the pattern-finding step needs a model you supply.

Is Recursive Human Intelligence Really Private?

The honest answer is local-first, not local-only — and the README is unusually candid about the split. This table is reconstructed from the README's own "What leaves your computer" section:

Each analysis call samples at most 80 user excerpts or 45,000 text characters, with obvious credentials redacted and session names replaced with anonymous codes before sending. The launch coverage frames this as "processing data already shared with AI assistants" — the excerpts going to your provider are cut-down pieces of chats you already sent to a model, not new exposure to a new party. That framing is fair as far as it goes, with one caveat the README states plainly: names and personal details in your messages can remain, so curate what the app is allowed to scan.

The part to configure deliberately is the background pass. Once you have connected a model and allowed sending, the Sunday 16:00 pass "sends excerpts to your provider without asking each time." If that rhythm does not suit you, turn it off in Settings → Background. Privacy here is a dial you set, not a promise you take on faith.

RHI vs Session Managers and Chat Search Tools

The chat-history tooling lane is crowded, so it helps to draw the lanes. RHI is the human productivity-feedback mirror: it reads what you typed and tells you how to type better. Session managers do something else entirely — arvre, for example, navigates your Claude Code sessions and their forks as a tree so you can resume and jump between them. RHI never manages sessions; arvre never judges your habits.

A third lane holds the session-search tools — agent-chat-reader, Agent History, Agent Logbook, Claude Trace, agent-trace-cli — which find, search, or resume past sessions, mostly as developer utilities or API plays. Those answer "where did I say that"; RHI answers "why do you keep saying it that way." The lanes are complementary, never rivals: navigation finds the session, search finds the message, RHI finds the pattern.

Who Is RHI For?

The primary audience is knowledge workers who live inside LLM assistants — and the launch coverage sketches three concrete personas. Developers using Claude Code for pair programming might discover they over-explain context in every session when a system prompt would suffice. Writers working with Codex could learn they fragment long-form content requests inefficiently. Product managers using Pi for brainstorming might find they circle similar topics without building on prior insights.

The common thread is volume: if you have months of AI chat habits piled up, you have the raw material RHI needs. If you chat with AI once a week, there are no patterns to find yet. The developer, Andrea Ritossa, reports already gaining insights into his own writing patterns through early testing — a sample size of one, but an honest one for a tool this young.

The Launch Story So Far

Recursive Human Intelligence was officially launched on November 3, 2026, via a Show HN post titled "Show HN: Recursive Human Intelligence: Actionable feedback from your LLM chats," linking to the open-source GitHub repo. The repo itself was created a day earlier, on November 2, and the first independent coverage — the emergent.sh launch piece — followed within about a day. As of November 5, 2026, the repo holds 13 commits and 4 releases, published under the Apache License 2.0, which the README summarizes as "free to use, change, and sell" with credit to Andrea Ritossa preserved in derived works.

It is early enough that the gaps are visible. There is no official docs site yet, which makes this page the independent recursive human intelligence explained reference the field currently lacks. Star count sits at 2 as of this writing — expect every number in this section to move. Also worth knowing what the app explicitly does not include: cloud sync, signed installers, arbitrary scripts, and an MCP server are all listed as not included, so plan your workflow around a single desktop.

Sources

Every fact in this article was verified from a source fetched on November 4–5, 2026:

• emergent.sh — "Recursive Human Intelligence Analyzes Your LLM Chat Habits" (launch coverage, ~November 4, 2026): app definition, how-it-works, privacy claims, developer quote, and the November 3 Show HN launch date.
• showhn.buzzing.cc — Show HN listing "Show HN: Recursive Human Intelligence: Actionable feedback from your LLM chats" (links to the GitHub repo).
• github.com/andrearitossa/recursive-human-intelligence — README: tagline, install commands, the three BYO-LLM providers and key handling, the "What leaves your computer" privacy table, the Sunday background pass, feedback examples, Apache-2.0 license terms, and build prerequisites.
• GitHub repository page — repo metadata as of November 5, 2026: created November 2, 2026; 13 commits; 2 stars; 4 releases/tags; Apache-2.0; Java Swing desktop app.

Frequently Asked Questions

Is Recursive Human Intelligence free?

The app itself is free and open-source under the Apache License 2.0 — free to use, change, and sell, with credit to Andrea Ritossa. But the analysis step is bring-your-own-LLM, so you pay your chosen provider (OpenAI, Anthropic, or OpenRouter) for the API calls the feedback runs consume. No key, no analysis.

Does RHI run AI agents or chat for me?

No. As noted above, RHI only analyzes chat logs you already created — it never runs agents, automates tasks, or chats on your behalf. If you want session automation, you are looking at a different lane of tools entirely.

Which chat platforms does RHI read?

Three: Claude Code, Codex, and Pi. It reads the logs already on your disk, so you need to actually use at least one of these platforms before RHI has anything to analyze.

Do I need Java installed to use it?

No. The install scripts bundle everything the app needs — "No Java needed" is stated explicitly in the README. You only need JDK 17 or newer if you want to build RHI from source or contribute to it.

Can I stop it from sending my chats to an LLM provider?

Yes. Excerpt sending requires your explicit consent when you connect a model, and the weekly Sunday background pass can be turned off in Settings → Background. Without a connected model key, nothing leaves your machine at all — but you also get no feedback, since the pattern analysis is the part that needs the LLM.