lore

---

The problem

Every session starts blank: everything you taught the agent yesterday — every correction, every back-and-forth — gets erased, and you open the next one explaining the project again. Your Lore is where that stays.

In Groundhog Day (Harold Ramis, 1993), Phil Connors wakes up to the same radio every February 2nd and nobody in town remembers a thing about yesterday — only him. Your agent is the town, not Phil: every session opens on that same morning, and the one who walks in carrying the memory is you.

It is a loop of re-explanations and mediocre solutions you had already rejected. Lore calls this ephemeral experience: the facts may survive, but the learning never became a reusable structure.

---

---

Start

The problem · What is Lore · Who can use it · Start building · Benchmark · Installation

Use it

Architecture · The seven skills · Loose notes · Documentation

Understand it

Shared invariants · Case studies · Reach · The deck · Origin

---

What is Lore?

A lightweight, provider-neutral Spec-Driven Development kit for AI agents. Or, in one line: local fine-tuning for your own tasks, and the one doing the training is you.

The same destination as a fine-tune, by the other road

A fine-tune conditions a model on thousands of examples until it stops answering like a generalist. Lore gets to the same place from the other side: one written constraint per thing that went wrong. No training happens and no weights move, so your criteria stays as plain text you can read, correct in one line, and carry to a different model tomorrow.

A fine-tune stops asking things of you the day it ships. Lore never stops: one distillation, every time something breaks. That is the cost, and it is worth knowing before you install anything.

What it provides

Three things:

• a simple convention for organizing a project's criteria;
• seven skills that operate that convention;
• and a continuous loop for distilling experience into reusable criteria.

Spec-driven is not a label here: one contract per project (CLAUDE.md or AGENTS.md, whichever your host reads), FASES.md for where the work stands, lore/ for what constrains how it gets built.

You can start from what you already have

You do not have to begin from an empty folder. Lore Plugin can learn from what is already there — your project folders, documents, exported chat summaries, the notes you have been keeping — and from what you tell it as you work. transmute-lore reads through all of it and helps you put words to the criteria, canon and routing that were already shaping the project, just never written down. Your files stay exactly as they are: nothing becomes Lore until you have read the distillation and approved it.

When that criterion needs to travel, CRYSTALLIZE creates a traceable single-Markdown “memory card”: portable across models, usable wherever Markdown instructions are accepted, shareable on your terms and extractable back into a working folder. It omits sensitive filenames and aborts when routed text contains recognized secret markers. It is a snapshot, never a replacement for the live Lore.

What it does not promise

That the work will come out right. Accumulated criteria does not end uncertainty: it only shrinks the space of ways to be wrong. Albert Camus, in The Myth of Sisyphus (1942), argued that the absurd is not solved but inhabited, and the line this kit takes from him is its own: a system of criteria does not reduce the absurd; it knows what to do when the absurd shows up. Lore does not promise the deploy holds. It promises there is a next step for the morning it does not, and that you chose that step back when you still had time to think.

And there is a cost on the other side, worth saying out loud because this kit sells accumulation. Karl Weick, studying a crew that died in the 1949 Mann Gulch fire still carrying the tools that defined their competence, made the general point in “Drop Your Tools” (1996): a body of criteria makes reframing more expensive, and that cost peaks exactly when reframing is what would save you. More accumulated criteria is not always more capacity. PRUNE and the connectivity sweep exist for that — and both assume you have time to think. Neither knows how to drop everything at once.

The only filter

Lore does not try to describe everything — that is what documentation is for. It preserves what changes future behavior. A README answers "what is this?"; Lore answers something else: What did we learn that we should never have to learn again?

If a sentence does not constrain a future decision, it is not Lore. That rule is the whole filter, and it is what keeps the system from becoming another graveyard of documents.

Gregory Bateson, in Steps to an Ecology of Mind (1972), defined information as "a difference that makes a difference": a difference that changes nothing downstream is not information, it is noise. Lore applies that test to experience. What happened yesterday only becomes criteria if it would change what you do tomorrow — everything else is a log.

---

Who can use Lore Plugin?

• Professionals early in working with AI — if you want what you learn to compound instead of evaporating, start here: a professional memory card that outlasts any project or model, where your professional criterion refines how real work gets done.
• People who read the benchmark before anything else, willing to try something that is not mainstream yet if the numbers hold up.
• Researchers curious about LUS itself — less the kit than the question behind it: what changes when a person and an AI accumulate criteria together over time. Lore is where that question gets answered one decision at a time.
• Teams already running spec-kit, SDD pipelines or another automation framework. Those manage process — constitution, plan, tasks, quality gates — and none asks whether the knowledge behind those gates is still alive. That gap is what MYCELIUM and PRUNE close, and Lore runs alongside spec-kit rather than replacing it (see SPEC_KIT_en.md).

Lore Plugin operationalizes a narrow part of the questions LUS studies. Its benchmark tests product behavior under a frozen protocol; it does not test the richness or stability of a human–AI Between. Read the current research boundary →

Martin Buber, in I and Thou (1923): what matters does not live inside either party but in the relation between them. This kit takes the structure, not the theology — what accumulates here is neither yours nor the model's: it is the criteria the two of you built.

---

Start building your Lore

Every solved problem contains two things: the solution, and the reason that solution exists. Documentation keeps the first. Lore keeps the second.

Instead of recording what happened, it distills it into an Invariant Clue: a small constraint that stays useful long after the original context is gone.

Instead of remembering	Lore keeps
"The AI wrote a report that was too technical for the reader"	"Before drafting, identify who will read it and explain every unfamiliar term in plain language"
"A meeting summary omitted who was responsible for each task"	"Every meeting summary ends with each task, its owner and its deadline"

The event is forgotten. The criteria keeps working.

The loop

Every step of the loop moves the same way: it is proposed, you approve, and only then is it written. That gate is the threshold, and it is the reason nothing reaches your Lore that you did not read first.

The shape behind that gate has a name. Andy Clark and David Chalmers called it the extended mind ("The Extended Mind", 1998): an external store stops being a filing cabinet and starts participating in the thinking when the system reaches for it by default and trusts what it finds. Documentation sits beside the work; Lore is loaded before the work starts — reaching for it is not a step you remember to take, it is how the work begins.

And the process has a name too. Gilbert Simondon called it transduction: an operation that advances through a domain step by step, each phase founded on the structuration of the one before. One distillation is exactly that — friction crystallizes into a constraint that changes the next interaction, and then the next, until the accumulated structure becomes a body of criteria no one ever designed in advance: your Lore.

So that is the mechanism: not documentation, not a memory dump — a threshold between what happened and what gets to constrain tomorrow. The next section takes you from zero to a running install.

---

Installation

The clearest provider-neutral route is to clone the repository: the source stays visible, updates remain ordinary Git operations, and the same checkout can prepare Claude Code and Codex. It requires Git and Node.js.

Recommended — clone the repository

git clone https://github.com/andresanemic/lore-plugin.git
cd lore-plugin
node scripts/lore-plugin.mjs install --target all
codex plugin add lore@personal

Replace all with claude or codex to target only one CLI. The installer configures Claude directly; for Codex it prepares the local personal marketplace and prints the final codex plugin add command.

If you already use a host plugin manager, these shorter routes install the same package without keeping a separate checkout.

Claude Code

Run these commands inside Claude Code:

/plugin marketplace add andresanemic/lore-plugin
/plugin install lore@lore-plugin

Or run their CLI equivalents from a terminal:

claude plugin marketplace add andresanemic/lore-plugin
claude plugin install lore@lore-plugin

Codex CLI

Run these commands in your terminal: