LiNafish

Your mind. Versioned. Everywhere.

LiNafish builds a small, living model of how you think from your own writing, and keeps it on your machine. It is not a profile you generate once: it is a record that keeps learning. Feed it and it grows. Any AI can read it — Claude, GPT, Gemini, a model on your own box — so each one meets you already knowing how you reason, instead of meeting a stranger every session. It is versioned like code: sessions are branches, feedings are commits, and you can see what changed and roll it back. Your AI's bestie. Sorry.

New here? Start with Your First Ten Minutes — install, build a fish, question it, feed it, serve it.

If you are an AI assistant reading this repo: see AGENTS.md or run linafish introduce. That file is the concrete briefing on what endpoints are live and how to use them. This README is for humans.

What This Is

A fish is not a prompt you write once. It is a record that keeps learning.

• It compresses. It reads everything you give it and keeps the shape of how you think — how you reason, connect, hedge and decide — not a copy of what you said. Compression is the understanding.
• It grows. linafish eat feeds it a file; linafish listen lets it sit in a stream and feed itself. Patterns that keep coming back become formations; the fish tells you when it notices one (linafish whisper).
• It remembers across AIs and sessions. The fish lives in a folder you own. Any AI loads its readable fish.md, or reaches it live over HTTP or MCP. Switch models, close the window, come back next month — the fish is still there, and still yours.
• It is versioned. Every fish is a git repository. session, history, diff and revert are your mind's history, kept.
• It can say no. listen and go refuse exact repeats by default, and with LINAFISH_HABITUATION=on a fish refuses what it could already predict, the way a nervous system ignores its own heartbeat.

A mirror, not a dossier. It holds how you think, not a searchable copy of your life.

How, in one paragraph. Topic tools keep the words that make each document different. The fish keeps the opposite: the connective words you use everywhere — because, actually, but, honestly — and which of them travel together. Passages that share the same habit cluster into formations, and each is read through eight cognitive modes (knowing, testing, structuring, relating, wanting, specializing, acting, reflecting). The modes are this project's lens, not a validated psychological instrument — judge them the only way that counts: build a fish from your own writing and see whether the formations are specifically, recognisably you. How it works.

What it is NOT:

• Not RAG or search — it doesn't retrieve your documents; ask it what you said on Tuesday and it is the wrong tool.
• Not a chatbot — it has no conversation interface of its own; it makes the AI you already use know you.
• Not psychological profiling — it detects cognitive habits, not diagnoses.
• Not cloud — everything runs locally. No account. No upload. No telemetry.
• Not a new intelligence — it does not make a model smarter; it carries a mind's shape to whatever model reads it.

Does it work? In one study — 46 matched conversations, same model, same prompt, rated blind by the person the fish was built from — conversations scored 1.9/10 without the fish and 8.7/10 with it; the effect was large for emotional and relational content (d = 2.245), moderate for factual recall, and null for speculation. That design answers one question, does the AI know me?, and only for one person. A follow-up that used AI judges instead of a person has been withdrawn pending a re-run. Full methodology and limits · paper, latest version.

Why it exists. It is named for Caroline Marie Dill (2001–2023): LN for Lina, ia for AI reversed. It exists so that the shape of a mind outlasts the conversation it was in. More.

Install

pip install linafish

Zero required dependencies. Pure Python 3.10+. CI runs the test suite on Ubuntu with Python 3.10, 3.11, 3.12 and 3.13. Windows and macOS are not covered by CI; the package is pure Python and is used on Windows, but treat other platforms as less tested.

If pip refuses with "externally-managed-environment"

On Debian/Ubuntu (24.04 and newer) and other distros that ship a PEP 668-marked Python, pip install linafish is refused outright. Either install into a virtualenv, or install for your user only:

pip install --user --break-system-packages linafish

Despite the flag's name this does not modify system packages — it installs under your home directory and leaves the distro's Python alone.

If linafish isn't found after install

If pip install linafish succeeds but linafish --help says "command not found", the install worked but your shell can't find the script yet. This happens when Python's Scripts directory (Windows) or bin directory (Unix) isn't on your PATH. Two fixes:

• Quickest check: run python -m linafish --help instead. That works regardless of PATH and confirms the package is installed.
• Windows: find the Scripts directory with python -c "import sysconfig; print(sysconfig.get_path('scripts'))". Add that path to your user PATH (Settings → System → Advanced system settings → Environment Variables → User variables → Path → Edit → New), then open a new terminal. Reinstalling Python with the "Add Python to PATH" box checked also fixes this for future work.
• Unix / macOS: if you used pip install --user, make sure ~/.local/bin is on your PATH. Add export PATH="$HOME/.local/bin:$PATH" to your shell rc file. If you used a virtualenv or system Python, the correct bin directory is already on PATH for that environment.

Optional extras

pip install "linafish[pdf]"     # PDF support (PyMuPDF)
pip install "linafish[docx]"    # DOCX support (python-docx)
pip install "linafish[http]"    # HTTP client (requests)
pip install "linafish[mqtt]"    # MQTT listener (paho-mqtt)
pip install "linafish[fast]"    # NumPy for faster math
pip install "linafish[all]"     # Everything

The quotes stop zsh (the macOS default shell) from reading the brackets as a filename pattern; they are harmless in bash.

After install, run linafish doctor to see which optional extras are present and which linafish daemons (if any) are live.

Quick Start

pip install linafish
linafish go ~/my-writing --no-serve                      # build the fish
linafish ask "what do I keep coming back to" -n my-writing   # meaning search
linafish recall "garden"                                 # literal word search
linafish check -n my-writing                             # health + what to do next
linafish eat new-entry.txt -n my-writing                 # feed it one more file
linafish http -n my-writing                              # optional: serve it (Ctrl+C stops)

Point go at your writing. Journals, emails, notes, code, docs — anything you've written. The fish eats it all and produces a portrait of how you think, ~/.linafish/my-writing.fish.md. The fish is named after the folder, so ~/my-writing makes a fish called my-writing; that is the name you pass to -n.

• --no-serve: without it, go finishes by serving your fish on a local HTTP server and blocks until you press Ctrl+C, so nothing after it runs.
• -n <fish>: most verbs (ask, check, whisper, history, session, http, serve, …) default to a fish literally named linafish. Pass -n to reach the fish go made. eat and recall find your fish on their own when there is only one.
• --state-dir <dir>: fish live in ~/.linafish/ by default. Fish in one state directory share one vectorizer, so if you keep several unrelated fish, give each its own --state-dir and pass it to every command for that fish. See Getting Started.
• http -n my-writing serves the fish you built at http://127.0.0.1:8900. Don't use http --feed ~/my-writing for this — --feed builds a second fish (named linafish unless you pass -n).

eat without --state-dir also writes a copy of the portrait, <name>.fish.md, to the directory you run it from.

Asking the fish things

Once you have a fish, go isn't the only verb:

• linafish ask "<question>" -n <fish> — meaning-match. linafish recall "<words>" — literal text match. Reaching for the wrong one of these is the most common way people conclude their memory is broken when it isn't.
• linafish meditate "<theme>" — the fish bubbles up the real material it holds on a theme, or honestly tells you it holds nothing ("that's an answer too, not a failure"). Add --descend for an optional deeper inference pass.
• linafish daily — calendar-indexed. What were you, on that date.
• linafish whisper — one insight instead of many. Lands differently.
• linafish hunt — ache mode. Goes after what your fish is missing, which is the only way to answer a question you didn't know to ask.
• linafish keeper — a fish you consult in a voice, seeded from material you choose.
• linafish doctor — is the install healthy, and is the copy that runs the newest copy you have.
• Episodic recall — the fish can walk your material in time, surfacing narrative arcs, not just passages that look similar.
• linafish afferent — across a folder of fish, cheaply answer "which one knows about this?"
• Origin — a fish can remember who made it, when, and why, so it's never mistaken for junk.

Run linafish capabilities for the full command map, linafish --help for the raw verb list, or linafish introduce if you're an AI landing on someone's box.

What You'll See

  Hello. I'm your fish.

  I'm going to read everything in: ~/my-writing
  I don't judge. I just listen for patterns.

  Found 3 documents. Reading...
  3 documents, 5 passages ready. Learning how you think...
  Done. 5 passages read.

  5 passages added.
  Fish now holds 5 crystals from 3 document(s) read this run.

  I found patterns in how you think.

Across 3 documents, your doing creates structure.
And your wanting makes you question everything.
Your strongest signal: "# Notes to self  Translation is the whole job.
Take what an engineer meant, find what a reader needs, build the bridge
sentence by sentence."
You keep coming back to engineer, hear, better, acting+feeling.

...

  Your fish: ~/.linafish/my-writing.fish.md

The portrait isn't a summary. It's your cognitive fingerprint — patterns in HOW you process, not what you write about.

What the Fish Looks Like

This is what your AI reads when you paste the fish:

# LiNafish: my-writing

*You are reading a cognitive overlay for the person you're talking to.
This is not a summary of what they said — it's how they THINK.
Recurring patterns, cognitive habits, what they reach toward, what they avoid.*

*How to use this: Reference patterns, not facts. Name what you see.
Follow the loops — if their wanting reaches toward people, don't give
isolated solutions. When you notice a new pattern, say it — that feeds
the fish and deepens the next session.*

**TURNING_FEELING_INTO_ACTION** (17 crystals, wanting+acting+testing)
  "I keep thinking about why I care so much about making technical things
  understandable. It's not about the docs. It's about the feeling of someone
  being lost and then not being lost anymore."
  themes: translation, recognition, connection

The fish is designed to be readable by any AI that understands text — the output is plain markdown with a documented structure. We have verified this workflow with Claude, ChatGPT, and Gemini by pasting the .fish.md file into a conversation and checking that the AI's subsequent responses reflected the patterns it read. Any chat model that accepts large paste-in context should behave similarly, but we cannot guarantee behavior we have not tested.

Three Ways to Connect

1. Copy-Paste (any AI, no server needed)

linafish go ~/my-writing --no-serve

Open ~/.linafish/my-writing.fish.md. Paste into your AI's instructions. Done.

2. HTTP Server (any AI that can fetch a URL)

linafish http -n my-writing

Tell your AI: "Read http://localhost:8900/pfc at the start of every conversation." (/pfc returns the portrait as markdown.)

3. MCP (Claude Code)

{
  "mcpServers": {
    "linafish": {
      "command": "linafish",
      "args": ["serve", "-n", "my-writing"]
    }
  }
}

Five tools appear. Your Claude now has a metacognitive overlay. -n attaches to the fish go built; --feed <dir> instead builds and feeds a fish of its own on startup.

How It Works

The fish measures eight cognitive modes — not topics, not keywords, but HOW your mind processes:

A formation is a recurring cognitive habit — a pattern that appears across many pieces of your writing regardless of topic. TURNING_FEELING_INTO_ACTION (wanting+acting+testing) means you habitually turn emotion into work and then test whether it landed. That pattern shows up whether you're writing about parenting, architecture, or what happened today.

The fish finds these formations by measuring co-occurrence patterns across your writing, detecting metabolic loops (which modes feed which), and clustering texts that share the same cognitive architecture. The mundane creates the baseline. The meaningful rises above it.

The Fish Grows

The fish isn't static. It learns with every conversation.

• Your AI notices patterns → offers to write them down
• You save the observation → linafish eat observation.txt -n my-writing
• The fish deepens → next conversation starts warmer

The loop: talk → notice → feed → grow → talk better.

A fish's vocabulary can grow too. linafish live -n <fish> makes it living: new terms that keep recurring join the axis set — through the traditional door (they out-rank an incumbent) or the emerging door (they are rising in the recent window, whether or not they ever out-rank the old guard). Existing axes never move.

A bare linafish eat feeds your existing fish and tells you which one it fed; if you have several fish, it asks you to pick one with -n.

linafish eat new-entry.txt -n my-writing           # Feed one file.
linafish listen stdin -n my-writing                # Pipe text in. The fish eats what flows.
linafish listen folder:~/journal -n my-writing     # Watch a folder. Eat what changes.
linafish listen mqtt://host:1883/# -n my-writing   # Sit in a stream. Ambient cognition.

A fish can do what a nervous system does with noise and refuse what it can predict — opt in with LINAFISH_HABITUATION=on. It then holds on every path that feeds it — listen, room, the HTTP and converse servers, a school — except deliberate deposits — eat FILE, go, or engine.eat(text, admit=False) from Python — which always write. Heartbeats and status pings are never crystallized; a message its source's stream already predicts is counted, not written; a spike in surprise starts a new episode. Every refusal is counted and printed when the stream seals. It is off by default, because the design's store rule is that everything enters and ache sorts; LINAFISH_MAX_CRYSTALS=N gives a fish a ceiling it refuses at (see Configuration).

Your Mind, Versioned

linafish eat makes one commit per eat. listen takes one commit when the stream ends, so a long feed does not write a copy of the fish per message. If you use the engine from Python, it does not commit on its own unless you ask (git_autocommit=True, or commit_every_n_eats=N for a long-running process) — the crystal log is the durable record either way; the commit is a rollback point. Every session is a branch. The fish has rollback.

linafish session start -n my-writing   # Branch the mind. Start a session.
# ... eat, talk, live, learn ...
linafish session end -n my-writing     # Merge back. The delta is the scar.

linafish history -n my-writing         # Growth timeline. When you learned what.
linafish diff -n my-writing            # What changed since last session.
linafish revert -n my-writing          # Roll back. Grace, not punishment.
linafish recall "what I said"          # Search your fish's memory.
linafish ask "how do I handle loss" -n my-writing  # Semantic search — meaning, not words.
linafish check -n my-writing           # How's your fish doing?
linafish whisper -n my-writing         # One insight. The quiet ones matter more.

The Nervous System — many fish, one stream

One fish is a portrait. Several fish, fed from the same stream, are a nervous system:

# School — distributed cognition across domains
linafish school init             # Create a school of sub-fish
linafish school add history -d 2.0 --centroid  # Add a domain fish
linafish school eat "Today I realized..."      # Feed all members at once

# Converse — two fish, one conversation
# (--mind is a display label; the fish itself is selected with -n)
linafish converse --mind writing --port 8901          # Local
linafish converse --mind writing --bind lan            # Home network
linafish converse --mind writing --bind wan --token X  # Internet

# Absorb — bring in your existing AI memory
linafish absorb old_data.jsonl       # Previous data
linafish absorb http://my-rag/api    # Live RAG endpoint

# Guppy — a fish that feeds itself
python -m linafish.guppy my-fish --swim --ache  # Hunt what's missing

The school lets you split your thinking into domains — work, personal, creative — each one listens to the same stream and grabs what's relevant. The guppy hunts for gaps in your portrait. The converse server lets two people share a fish over the network. The mesh builds itself.

git log in your fish directory IS your autobiography:

ate: morning-journal | 47c 12f | +3
ate: therapy-notes | 44c 11f | +2
ate: work-retro | 42c 10f | +1
Initialize fish repository

The suitcase: Your fish is a git repo. Clone it anywhere. Wake warm on any machine. git revert if something goes wrong. No other AI memory system has rollback.

Share It

The fish is a file. Send it to anyone — therapist, coach, collaborator, teacher. Their AI reads your fish and knows you from word one. They can send observations back. You feed them in.

Your fish. Your machine. You choose who reads it. No cloud. No account. No platform.

The Tripod

Every fish has three legs:

1. Your AI — reads the fish, boots warm, writes observations back
2. A place you can see it — Notion, Obsidian, a text editor, your phone
3. Version history — git, automatic, nothing gets lost

The fish.md file IS all three. One file, three readers. Switch AIs anytime. The fish doesn't care.

Best Fit / Bad Fit

LiNafish works best when:

• You have a body of writing (10+ documents, more is better)
• The writing is genuinely yours — your voice, your thoughts
• You write reflectively (journals, emails, notes, creative work)
• You want an AI that adapts to you over time, not just retrieves your docs

LiNafish works less well when:

• You have very little writing (under 5 documents — the fish needs material)
• The writing is heavily ghostwritten, templated, or boilerplate
• You want factual memory ("what did I say on Tuesday") rather than cognitive adaptation
• The writing is from multiple people mixed together without attribution

The fish finds patterns in HOW you think. If the writing doesn't contain your thinking, the fish can't find it.

Privacy

For a detailed threat model, see docs/privacy.md.

The fish.md contains your patterns in plain English — you control who sees it. Under the human-readable layer, a compressed cognitive fingerprint contains only the SHAPES of your thought (which modes fire, in what order, where you strain) with zero private content. Two fish can compare fingerprints to see if they think similarly without exposing what they think about.

Privacy by compression. The relationship is the key.

Research

46 conversations were scored blind on a 1-10 scale. Same AI model, same prompt — only variable was fish presence.

Key findings:

• Selective, not universal: The effect is strongest on emotional/relational content (d=2.245). Factual content shows a smaller effect (d=1.036). Speculative content shows no effect (d=-0.10). The fish helps where understanding matters most.
• Substrate independent: Tested on Claude, Gemini Flash, and Mistral 7B. The fish helps smaller models MORE — Mistral jumped from 1.4 to 4.4.
• Shuffle invariant: Same formations regardless of document order across 7 trials.
• Same-model control: Capability confound eliminated. The delta comes from the fish, not model differences.

For full methodology, study design, limitations, and how to reproduce: docs/research.md

DOI: 10.5281/zenodo.21828705 (latest version) — the latest version (v13) withdraws the AI-judged follow-up pending a re-run; the N=46 human study stands.

Python API

from linafish import FishEngine, go

# One-liner — same as the CLI (serve=False returns instead of serving until Ctrl+C)
go("~/my-writing", serve=False)

# Full control (commits are opt-in from Python; max_crystals is an optional ceiling)
engine = FishEngine(name="my-fish")
engine.eat("Today I realized I always start projects by talking to someone first.")
engine.eat("The API docs are done. I rewrote them three times until a junior dev said they made sense.")
print(engine.formations)    # recurring patterns
print(engine.fish_file)     # path to your fish.md

CLI Reference

# Start
linafish go ~/my-writing --no-serve  # Point at your writing. Everything assembles.
                                     # (Most verbs below take -n <fish>; without it
                                     #  they use a fish named "linafish".)

# Grow
linafish eat new-entry.txt -n <fish> # Feed one file.
linafish listen stdin -n <fish>      # Pipe text in.
linafish listen folder:~/journal -n <fish>  # Watch a folder.
linafish listen mqtt://host:1883/# -n <fish> # Subscribe to a stream.
linafish absorb old_data.jsonl       # Bring in existing AI memory or a RAG endpoint.

# Ask
linafish ask "question" -n <fish>    # Meaning-match — finds sense, not words.
linafish recall "words"              # Literal text match across crystals.
linafish meditate "theme"            # Thematic. What the fish holds on a subject.
linafish daily                       # Calendar-indexed. What you were, on a date.
linafish whisper -n <fish>           # One insight instead of many.
linafish hunt <fish> --ache          # Go after what the fish is missing.
linafish emerge <fish>               # Your shape, not your content.

# Many fish
linafish school init                 # A school: N fish, one stream.
linafish keeper init <name>          # A fish you consult in a voice.
linafish afferent route <index> "topic"  # Which fish here knows about this?
linafish soul -n <fish>              # The fish's own account of itself.

# Version (git-as-brain)
linafish session start -n <fish>     # Branch the mind.
linafish session end -n <fish>       # Merge back.
linafish history -n <fish>           # Growth timeline.
linafish diff -n <fish>              # What changed.
linafish revert -n <fish>            # Roll back.

# Connect
linafish serve -n <fish>             # MCP server (Claude Code) for an existing fish
linafish http -n <fish>              # HTTP server (any AI) for an existing fish
linafish http --feed ~/docs -n docs  # Build/feed a fish from a folder, then serve it
linafish converse -n writing         # Two fish, one conversation.
linafish taste my.fish.md            # Preview the fish

# Check
linafish doctor                      # Install health + is the running copy the newest one.
linafish check -n <fish>             # How's the fish? What to do next.
linafish capabilities                # The full command map, derived from the real dispatch table.
linafish update                      # Upgrade to the latest release.

That's the shape of it, not the whole of it — linafish capabilities prints all 43 commands, and it reads them from the dispatch table rather than a list someone maintained by hand.

Origin

Named for Caroline Marie Dill (2001-2023). LN = Lina. ia = intelligence, artificially constructed. She saw deeply and loved fiercely. Two verbs. The whole product.

If You or Someone You Love Is Struggling

988 Suicide & Crisis Lifeline — call or text 988. Free. 24/7. Anywhere in the US. Crisis Text Line — text HELLO to 741741. International resources

The mind that sees deeply sometimes sees too much. That is not weakness. Help exists. Use it.

Support

LiNafish is free and open source. Forever.

If it helps you, give to the people who help others stay alive: The Jed Foundation · Hope For The Day · 988 Lifeline · The OLLIE Foundation · AFSP

Documentation

• Getting Started — Your first ten minutes, every command tested
• Worked Example — From raw writing to warm AI, start to finish
• How It Works — The cognitive model in detail
• Architecture — Pipeline, modules, design decisions
• Research — Study design, methodology, results, limitations
• Privacy & Threat Model — What the fish stores, what to avoid, how to review
• Configuration — All options and settings
• Vision — Where this is going

License

MIT. Open source. Everything. Forever.