The AI Memory Bible

THOR - a memory for your AI coding assistant

Most memory tools can only remind your assistant. THOR can refuse.

Give it a note with a proof attached - a small check it can run right now, like "this file still contains that line." From then on the note stops being advice. THOR refuses the matching edit or command before it runs. Not a suggestion your assistant is free to skip. A hard stop.

That proof runs live, at the exact moment it matters, against your actual project. A note that has gone stale, that no longer describes what is really there, quietly stops counting as proof - so it can never block you by mistake.

Other local memories can refuse a write too. What is different is where the block comes from: a proof, checked live, never a rule taken on faith. And the whole loop that produces it, learning included, runs in one place: this machine.

Most notes only inform: pinned so they are handed over at the start of every session, or surfaced the first time your assistant reaches for the file or command they are about.

A note that keeps being served without ever being judged makes THOR stop the turn and ask whether it helped or got in the way. Two bad verdicts in a row retire a note from being served again, though it stays there if you go looking for it. That ask only ever names a note your own session actually saw fire - never one that only ever showed up during other, unrelated work. Answer for a note once and this session leaves it alone for good, no matter how many more times it fires before you are done.

A health check names what has rotted, out loud: proofs that now come out false, notes pointing at nothing, notes too crowded to ever be shown.

Everything that is not code - a recipe, a book, a training log - lives in its own library, kept apart, so it never competes with a deploy rule for room.

Everything stays on your machine: no account, no key, no cloud - including what it learns about which of your notes are worth keeping.

THOR is free and GPLv3. If it earns its keep, you can support the work: buy me a Ko-fi or become a YouTube member - no pressure, no paywall, everything stays open either way.

Are you an AI assistant, setting THOR up for the person you work with? Go straight to AGENTS.md. It is written for you.

Prefer watching? The video above walks through what THOR is and why it exists.

What it does

Remembers what you tell it. A rule, a gotcha, a decision, the shape of the project. Once, in your own words. It stays until you change it.

Hands it back at the right moment. Not a search box you have to remember. The note arrives while you are touching the file or running the command it is about, in a conversation that never heard it.

Stops a wrong change, not just warns about it. A note carrying something checkable can refuse the write outright. Most notes only inform, and that is deliberate: a rule that blocks honest work is the most expensive thing this system can do.

Keeps projects apart. Every project has its own memory. One repo's rules never leak into another.

Says when it has rotted. It counts its own dead ends: notes pointing at files that moved, notes nothing ever reads, notes crowded out by louder ones. Out loud, in plain language, so you can fix them.

Stays on your machine. No account, no key, no server. Nothing is sent anywhere, ever.

What that looks like in practice

Months ago you found out the hard way that this project is pinned to an older Node, and anything newer breaks the build. You said so once, and moved on.

Today a fresh conversation opens package.json to add a dependency, sees an engine range that looks out of date, and is one helpful edit away from bumping it. Right then, before it types, your own sentence is in front of it.

That is the whole idea. Not a search box you remember to use. A memory that shows up on time.

"I already have a CLAUDE.md for that"

Most assistants read a rules file at startup - CLAUDE.md, AGENTS.md, .cursorrules. It helps, and it runs out of road quickly.

Past a certain size it becomes a phone book, and nobody reads a phone book front to back. Your assistant skims it, takes the gist and moves on. The rule was in there. It got skipped. Nothing looks wrong afterwards, because the line is still sitting in the file, so you go on believing you are covered.

Here is the part that is genuinely different. Picture a fresh agent, no history, no idea what this project has already cost you, one keystroke away from the exact write that broke production last spring. A rules file would have mentioned it somewhere on page four. THOR stops the keystroke. The write does not happen - and what stops it is the note you wrote, the day it broke.

That is the whole promise: not better advice, but a wrong change that does not land.

Getting there is not free, and it is worth knowing before you start. A note only earns that power if it is written to earn it: tied to a real file or command, carrying something checkable that shows it still applies. THOR ships with a handful of starting notes that teach exactly that, and refuses the ones that cannot work. AGENTS.md spells out the rules of the game in full.

Three things make that work, and they all live in one file on your machine:

1. Nothing is ever lost. Every note is kept forever. Change your mind and the old version stays too, so you can always look back at what you used to think and when it changed. If two versions of a note ever conflict, THOR keeps both and tells you, rather than quietly picking one and throwing the other away. It is the same care you would give your source code, given to the things you know.

2. It arrives at the right moment. THOR checks your memory on every message you send. And the first time your assistant reaches for a file or runs a command that one of your rules is about, that rule gets put in front of it right then. Before the mistake, not after.

3. Your assistant looks after it. It is not a notebook you have to fill in by hand. Your assistant can add notes, correct them, retire ones that stopped being true, and flag which ones actually helped. A THOR that is used well is a THOR your assistant is quietly tidying as you work.

It reads your code too. Point it at a project and it takes in the source and the documentation, so "how does this bit work here" gets answered from your actual project instead of a guess.

Everything stays on your machine. No cloud, no account, no subscription, and nothing to sign up for. If some optional piece is missing, THOR quietly falls back to a simpler way of working instead of breaking.

A note that can actually stop you

Showing a warning at the right moment is worth a lot, and for a long time that was all THOR could do. A note could speak. It could not refuse.

Version 2 lets a note carry a proof of its own currency: a small check THOR can run right now to see whether the note is still true of your project. "This file still contains that line." "That file is still there." "This character never appears in anything we write." Or, for catching something left out rather than something wrong: "every agent I spawn names which model to do the work with."

That changes what a note is allowed to do:

• A note whose proof runs and holds right now may stop a wrong change outright.
• A note backed by words alone may warn, and only warn. It can inform your assistant; it can never forbid.
• If the proof cannot run - the file moved, the path is gone - nothing is blocked. It is reported as needing a look.

That first kind of stop reaches further than an edit made through your assistant's own tools: deleting the protected file, emptying it out, or overwriting it from a command your assistant runs counts as the same wrong change, and is stopped the same way.

The reason for the split is uncomfortable and worth saying out loud. Notes rot. You write one, the project moves on, and the note quietly becomes wrong. A tool that let any old note block your work would spend most of its time blocking you for reasons that stopped being true months ago. So THOR only hands that power to notes that can prove, at that exact second, that they still describe your project.

Which is why, as the top of this page already said, most of your notes will never block anything. The health check prints how many can, and you should look at it. On the author's own memory, when this was first measured, 2 notes out of 2999 could prove themselves; a day of deliberate work took that to 256. It moves by hand, because deciding what proves a note is a judgement about that one note.

That the number is printed at all is the point: a safety net nothing is attached to looks exactly like a safety net that works.

THOR asks, so you do not have to remember to

Leaving that to whoever thinks of it means it never happens. So THOR asks, by itself, in two places.

When a note is written. A note you call expensive, or one that spells out a command, a flag or a filename, is not stored until one question is answered: is there a text whose presence means the mistake is happening? If there is, the note gets a proof built on exactly that text. If there is not - and often there is not, because "check with me first" has nothing to catch - you say so and the note goes in unchanged. Both are real answers. Only saying nothing is not.

For the notes you already had. Once per session, THOR picks one note that names something concrete, has never been asked, and holds the turn until it is. One at a time, forever, so a memory written before any of this existed still gets worked through instead of being declared hopeless.

A caution worth stating plainly: THOR can prove that a note is wired so a matching change would be stopped. It cannot know whether the text you typed is the text the real command uses. A misspelled fragment is wired perfectly and guards nothing. That is why the health check reports two different numbers - how many notes could refuse something, and how many ever actually did. Trust the second one.

Asking about one note at a time is a brake, not a broom: it keeps the pile from growing while you work, not a way to clear it in one sitting. For that, run /thor-eval at the end of a session. It walks through everything the memory has been served without a verdict, repairs what has quietly drifted out of date, and settles the debt in one pass instead of one note a day. A session that never runs it simply leaves that debt for the next one to inherit.

A memory with a spine

Version 1 remembered well and never argued. It would hand your assistant a note at the right moment and hope. Version 2 is the same memory with a spine.

• A note can refuse. The headline, and the rest of this section is about it. Version 1 could only speak.
• Bad notes no longer get in. A note that cannot ever fire, has nothing that would prove it wrong, runs on for a page, or simply repeats one you already have: turned away at the door, with the reason and the fix. Version 1 stored whatever it was given, and you found out months later that half of it was unreachable.
• It counts what it actually did. How many notes can refuse something, how many ever have, how much of your memory nothing re-reads. Version 1 had no number for the one thing it was built to do, which is how a safety net stays broken for a year.
• Maintenance is no longer optional. At the end of a turn it asks for one thing: judge a note that keeps firing, fix one you just filed where it will never be seen, answer whether an expensive note can refuse anything. One at a time, and it will not be waved off with a promise to do it later.
• It costs almost nothing to carry. A note is capped at 300 characters and a block at four notes and 1200 characters, so what lands in the conversation is a few hundred tokens, not a whole rules file re-read on every turn. A memory of ten thousand notes costs the same per turn as one of fifty. Nothing calls out to a model to decide what to send - it is a local program reading a local file, so there is no network round trip in front of your keystroke. The one exception is the handful you deliberately pin: those are read out in full at the start of a conversation, so pin sparingly and the cap does the rest. Measured on the author's own machine and memory: about 110 ms per call and around 1500 characters delivered when something applies, against roughly 660 ms for version 1. One machine, one memory - treat it as an order of magnitude, not a specification.
• Stale notes are hunted, not left to rot. A note whose proof comes out false is reported instead of quietly going on being wrong. A note that keeps firing without anyone ever saying whether it belonged gets asked about, and two verdicts of "this did not belong" retire it from every channel while leaving it findable. In version 1 a note that went wrong simply stayed.
• Crowding is visible and refused. Only a few notes fit in a block, so notes compete. Version 2 counts that competition, tells you when you have just stored something onto a spot too crowded to ever show it, names what is holding the place, and refuses the write outright when every spot the note could take is already full of heavier ones. A note tied to one file, folder or command is refused the moment that single place already holds as many as it can ever show - lighter rivals count too, so a heavier note can no longer bump a lighter one out of sight unnoticed. Version 1 accepted it and said nothing, which is how a memory fills up with advice nobody will ever see.

A second memory, for everything that is not code

The memory above is built for work. It has a gate, notes that interrupt you, and a hard cap on how much ever reaches the conversation - all of which is exactly wrong for a recipe.

So 2.1 adds a library, and it is a genuinely separate thing: its own file, its own two commands, and no way to reach the first memory at all. Nothing you put in it can ever interrupt you, compete with a note, or take up room in a block. You only ever see it because you asked.

It works the way a shelf works.

• Everything lives on a shelf, and shelves do not nest. Books, recipes, a training log, what you spent. Filing something without naming a shelf is refused, and the refusal lists the shelves you have, so your assistant picks from real ones instead of inventing a name.
• Only you create a shelf. If nothing fits, your assistant has to ask you what the new one should be called - and once you answer, it repeats your own word back and the shelf is created, with the entry filed onto it, in that same reply. It can never invent a name or skip the question. This is the rule that stops a tidy list of eight from becoming a sprawl of sixty.
• A shelf that grows gets labels, never a split. Two hundred recipes on one shelf, filtered by "bbq" or "dessert", stays one shelf. That is what keeps the list of shelves short enough to hold in your head.
• You get an index, not a wall of text. Open a shelf and you see one line per entry. Ask for one by number to read it whole.
• The same thing twice is refused, pointing at the entry you already have.
• Nothing is ever deleted. Your assistant can retire an entry that no longer belongs - it comes out of the listing and out of search, and stays fully readable by its own number for as long as the library exists.
• A search never answers "nothing". If your words miss - and they will, since the words you ask with are rarely the words you wrote - it hands you the shelf to read instead. Asking for "ribbetjes" when you wrote "ribben" finds it.

Getting started

The short way - one command. It downloads the latest release, checks the download against the checksum published next to it, unpacks it into your home folder, and runs the whole setup.

Windows, in PowerShell:

irm https://raw.githubusercontent.com/nworks3d/THOR-memory-bible-for-AI/main/install.ps1 | iex

Linux:

curl -fsSL https://raw.githubusercontent.com/nworks3d/THOR-memory-bible-for-AI/main/install.sh | sh

No administrator rights, and nothing is installed outside your own user folder. The automated setup targets Claude Code today. It touches two of its files - Claude Code's own settings and the list of tools it may use - and backs up both before it does. Rather read the script before you run it? Open that same link in a browser first. There is no macOS build yet, so on a Mac take the route below. Using a different assistant? See Using it from another assistant below.

Or build it yourself. You need a Rust toolchain. Nothing else: no key to get, no model to download first, no account.

cd thor2 && cargo build --release --features semantic

--features semantic is not optional, and leaving it off fails silently. Without it everything still builds, still runs, and still answers every word-for-word search correctly. What stops working is searching by meaning: it returns nothing at all, with no error anywhere. The reliable way to tell the two apart is size. Look at thor2/target/release/serve.exe - over 20 MB is the right build, a few MB is the wrong one. Build it again with the flag.

Then run the setup yourself. That is the same step the one-command install ends with, and in the normal case you type no paths at all:

thor2/target/release/install.exe

It finds your assistant's own two configuration files by itself, creates your memory if you do not have one yet, wires THOR into your assistant, and registers the part your assistant writes through. It also installs a shared set of git hooks that run in every repository on this machine, which keeps the reading of your code fresh after each commit, and each one hands control back to your repository's own hooks so nothing already there stops working. It backs up both files before it touches them, it never removes anything it did not put there, and running it twice changes nothing the second time.

A brand new memory does not arrive empty. It gets a handful of short notes on how to write a note that comes back to you later, and your assistant is handed them at the start of every conversation from then on. That matters more than it sounds: an assistant with nothing in front of it writes notes in a shape that never fires again, and neither of you would notice for weeks. They are ordinary notes - unpin one, rewrite it in your own words, or throw it out. An existing memory is never seeded, so upgrading never pushes anything into your own notes.

You only reach for a flag if your setup is unusual: --settings and --mcp-json send it at other files, --db and --serve-exe override where it looks, and --no-mcp sets up a memory your assistant can read but not write, on purpose. If a program it needs is missing, it stops and says so rather than installing something that would sit silent.

Then restart your assistant. The part it writes through only comes alive after a restart. Until then, it can already read the memory but not add to it.

Step 3 - check it.

thor2/target/release/doctor.exe --db "C:\Users\you\AppData\Local\thor2\thor.db"

Sixteen plain-language lines, one per part: which build you are running, whether your memory is healthy, whether searching by meaning is switched on, how many of your notes can prove themselves, how many point at files that are no longer there, and how many are bound to something that can never happen. It changes nothing.

To see just the build number - useful when you are reporting a problem and need to say exactly which version you have - run doctor as above and read its first line, or run any of the programs in thor2/target/release/ with --version, for example thor2/target/release/serve.exe --version.

One of those lines only speaks up when it has something to report: if it ever says your memory's own log file has outgrown the memory itself, something is stopping a save from ever finishing - close any other program that might have that same memory open, and run the check again. Newer builds also cap how big that log file is allowed to grow, so this should be rare.

Step 4 - give each project its own memory. From that project's folder:

thor2/target/release/install.exe --project "my-project"

This matters more than it sounds. Skip it and THOR gets worse the more you use it, because every search starts competing with projects you were not asking about. All it does is write a small file called .thor-project holding that name, so you can also just create that file yourself. It refuses to change a name that is already there, because renaming a project's scope would strand every note already filed under the old one.

If you are the assistant doing the setup, AGENTS.md is the walkthrough for the steps above.

Using it from another assistant

The automated setup above is for Claude Code specifically. Any other assistant that can be pointed at an external tool server can still use THOR, through the container published alongside every release instead of a local build.

A generic client config, the shape most tool-calling assistants expect:

{
  "mcpServers": {
    "thor": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-v", "thor-data:/data",
        "ghcr.io/nworks3d/thor-mcp:2.4.1"
      ]
    }
  }
}

Run it as a container

docker run -i --rm -v thor-data:/data ghcr.io/nworks3d/thor-mcp:2.4.1

On an empty volume, the container creates a memory and seeds it with THOR's own handful of starting notes before it starts answering - the same first step the setup above takes. Point -v at a folder or named volume that already holds a thor.db instead (a copy of the one your own build created, say), and that one is opened untouched. -i keeps input open, which is what talking to it over stdio needs; drop --rm if you would rather keep the stopped container around than have it clean up after itself.

Stay in one conversation

The old advice was to start a fresh chat often, because long ones got worse and you lost everything anyway. With THOR that advice is out of date. One long conversation is now the better habit.

When a conversation gets long, the assistant's tools squeeze out the older parts to make room. THOR covers that moment: your standing rules come straight back, and it nudges your assistant to write down anything important that was never saved. Starting a fresh chat is covered too - your rules and your project's background are loaded in from the start.

So stay in one conversation while you are on one piece of work. Start a fresh one on purpose - because you have moved on to something else, or because this one has talked itself into a corner - not because it is getting long.

What the first week actually looks like

Worth knowing before you start, because the beginning is the least impressive part and it is easy to conclude too early that nothing is happening.

Day one, it stops nothing. A fresh memory holds a handful of starting notes: how to write one that comes back, plus the honesty and agent-spawning habits that hold on any project. The part of THOR that can refuse a wrong change only works on notes that carry a proof, and you have not written any yet. So on the first day you get those notes at the start of a conversation and a nudge at the end, and no refusals at all. That is not a fault; there is simply nothing yet to refuse with.

Your first note will probably be turned down. It asks for two things most people leave out: when the note should come back to you, and what would show it had gone wrong. If the note is about something expensive, or names a command or a filename, it asks a third: is there a text whose presence means the mistake is happening? Answering "no, there is nothing to catch here" is enough, and often it is the truthful answer. The refusal names everything that is missing at once and says what to write instead, so the second attempt usually lands. It is strict on purpose - a note nobody can ever prove wrong is a note that quietly stops being true.

The value arrives once you have notes about real places. A note tied to a file, a folder or a command comes back exactly when you touch that thing. A handful of those is worth more than fifty general ones, and after a week or two of writing them down as you go, your assistant stops asking you the same questions.

Then check what you have built with doctor. It tells you plainly how much of your memory can actually stop a wrong change, how often it has, and which parts nothing ever re-reads.

If you arrive with a memory you already have

Everything above describes a memory that starts empty. If you are coming from version 1, or from any pile of notes written before proofs existed, the shape is different and worth saying plainly, because the obvious plan does not work.

THOR asks you about one old note per session. That is a brake, not a broom. It keeps the pile from growing while you work, and it was never meant to clear one: at one a day, a backlog of thousands outlives you.

The broom is the health check, pointed at the folder your projects actually live in:

doctor --db <your thor.db> --checkouts <the folder holding your projects> --full

That names every note whose anchor points at nothing, every proof that now comes out false, and every note stored somewhere too crowded to ever be shown. Set aside an afternoon rather than a coffee, and go through it in one sitting. The list is long because the memory is old, not because anything is broken.

Two things make that afternoon safe to be decisive in. Correcting a note keeps the old version, so nothing you wrote is lost and you can always read back what it used to say. And removing one does not delete it either: it stops being handed to anyone, stays findable, and the reason you gave stays attached to it.

Does it work?

Use it for a week and see whether your assistant stops asking you the same things. That is the only test that answers the question you actually have.

THOR was measured head to head against another memory tool for months, and those numbers are not here any more. Not because they were bad - they were good - but because a score measured on someone else's notes tells you about them, not about you. The tool is here. The verdict is yours.

The sixteen tools

What your assistant actually calls, grouped the same way the memory itself is split in two.

Code lane - projects, rules, how to work:

• remember - store a new rule, note, report or lookup entry
• revise - correct one that already exists, instead of a near-duplicate; taking teeth away from a rule that can already stop a mistake - dropping its check, lowering its severity, or narrowing where it applies - needs a reason too, kept in its history
• retract - remove one that turned out wrong (a reason is required; nothing is deleted)
• resolve - settle two versions of the same fact that diverged apart
• pin - make an existing note a standing rule, served at every session start
• unpin - stop it being one
• get - show one item whole, by id
• lookup - search everything - every project, by scope, by key, or by free text
• history - walk one item's whole past, oldest first
• mark - judge whether something you were served actually helped
• status - how much of the memory can actually stop a mistake, and how much has gone stale
• search_code - search the project's own indexed source code
• where_used - find every place one function, struct or variable is defined and used
• outline - see everything one file declares, in line order

Library lane - recipes, books, training, expenses, kept apart from the code lane:

• library - read a shelf's index, one entry, or search across shelves
• shelve - file a new entry, correct one that already exists, or retire one

Documentation

Thanks

• MakerViking - for the inspiration and the great fight. This project would not exist without the spark, and it would not be half as good without someone worth pushing against. Skål!
• mimir (MakerViking/mimir) - the reason THOR exists at all. In the old Norse stories, Mimir guards the well of knowledge; here it set the bar THOR had to clear, and for a long stretch it cleared plenty of its own. Every early comparison in this project was against mimir, wins and losses both published on purpose, because a rival that good deserves honest numbers.
• Ideas borrowed, both ways. Two things THOR does came from mimir's own work and were rebuilt here in THOR's own way, and mimir in turn credits THOR for reading code into memory and for checking memory on every message - exactly the kind of exchange open source is for. Thanks, MakerViking.

Support this project

THOR is built by N-Works 3D. If it has earned its keep - saved you an explanation, caught a mistake before it cost you, or just meant you did not have to start from scratch - there are two easy ways to help keep it going:

• Buy me a Ko-fi - a one-off, whenever you feel like it.
• Become a YouTube member - monthly, if you want to keep it going.

No pressure and no paywall - it all stays open either way. Skål, and thanks for reading this far.

Contributing

Bug reports and pull requests welcome. THOR is a memory your assistant is supposed to trust, so the bar is being right rather than having more features. The checklist is in CONTRIBUTING.md.

License

GPLv3.