HTML is all you need.
Near-limitless AI memory, kept in a plain local HTML file you own.
An addressable wiki that remembers it all.
The fix for context rot — across models, across apps, across years.
A doc.html is one ordinary HTML file — the plain kind people have been making since the early 90s. The same HTML happens to work for memory and instruction an AI agent or human can read, navigate, and update. Markdown is a low-bandwidth bridge that can only go so far. The file's manifest serves as a summary a model can retain in context, the way a skills.md or agents.md file would. One file can outgrow any context window, but a model or human can still retain all of the knowledge. Text, media — anything that can be shared with HTML can serve as addressable memory you can navigate, retrieve, update, and even share.
no server. no account. no javascript. open source
Today the words are tokens, and there are only so many tokens an AI model can ingest before succumbing to context rot and hallucinations. There is a way to keep the knowledge without losing the thread. doc.html.
three bodies, one formatone file: the map on top, the pinned sections below
the document you can queryOne file is a document with its index up front. Many files, linked, are a corpus: combine them, and a model can retrieve from the set the way it would from a database or a RAG index — with no database running and no index built. It pulls the sections a question needs and cites them by id and hash. It works; the receipts wall below holds the score.
see the map — the smallest one »corrections supersede; history stays addressable
the memory you keepThis is what your agents.md or CLAUDE.md wants to grow into: the decisions, conventions and gotchas of your project, kept long-term in one file any model can open. When a decision turns out wrong you append the correction; it supersedes, and the old entry stays, still saying why it made sense at the time. Next year's model opens the same file and already knows what you decided today.
start your memory file »turn 0042 by one model, turn 0044 by another — same conversation
the chat that continuesnewest bodyContext rot hits conversations first: the longer the thread runs, the less of it the model actually holds. A chat body writes the conversation down as it happens — every turn addressable, finished turns sealed with a hash — so a model can rejoin a conversation it was never in, and the thread outlives the app it happened in.
continue a conversation »
what the trinity buildseverything below is the same three bodies, put to work. the wiki is the memory body on a shelf.
the model walks the links like a librarian, not a vacuum
the wiki that remembersA wiki of doc.html pages, linked like any other website. A model checks the index, pulls the two pages the question needs, and leaves the rest on the shelf. A human can open those same pages in a browser and see an ordinary wiki page — because it is one. Our wiki also serves as documentation for the format — read more here.
walk the real wiki »see it in the browser window »
on the shelf right now — six essays, each its own pinned page:
the steer core is always read; everything else waits behind the map
the agent's inheritanceThe memory body with a job. agents.html is this repository's own memory: instructions, decisions, and unfinished work in one file every agent opens on arrival. A short steer core is read every time; the rest waits behind the manifest until a task names it. We run the project on it — including the section that lists what has not been demonstrated.
read the repo's own memory »
how it works — give the model a map, not a pileread the manifest — a few KB, even for a huge document.
name the two or three sections the question actually needs.
load exactly those bytes — nothing else enters context.
run the reader. it recomputes each section's SHA-256 and names any section that disagrees.
<!-- this is most of the format --> <nav id="manifest"> <a href="#fold" data-witness="771feff36…" data-char-count="75"> <span class="title">The Fold</span> <span class="summary">Append-only memory; correction is supersession.</span> </a> </nav> <section id="fold" data-witness="771feff36…"> <h1>The Fold</h1> <p>Memory is append-only; correction is supersession.</p> </section>
try it in 60 secondsno install, no account.
Read this document using the orientation it provides. Tell me what sections are available, then answer my question using only the relevant sections.
git clone https://github.com/NDOTO-G/doc.html cd doc.html node tools/verify.mjs examples/memory.doc.html
the receipts wallsealed first, said second. where a number came from a sealed run, the tile takes you to it.
| the number | what it proves | the receipt |
|---|---|---|
| 72.5 MB | one sealed file, far past any context window — a model read its way through and cited its way back | apparatus-memori · PR #48 |
| 240/240 | every citation checked against the actual bytes. none invented | apparatus-memori · PR #48 |
| 1 byte | change a single character and the next check catches it | the record — integrity |
| 4 langs | four readers, four languages, one written spec — a blind implementation matched every expected verdict; one raw-text cell split the four | reader-convergence · PR #72 |
| 2 readers | a forged witness fooled one reader — the other refused. the forgiving parser was deleted, not patched, and the pair re-sealed as a unit | reader repair · sealed |
| 0 | vendors who can take this away from you | public domain · no owner |
a homepage holds the short rows — the evidence page carries the rest, losses included.
[ ↑ top ]
a format that names only its strengths leaves you to guess at the rest. these are the limits, each with its why and its answer.
each limit, with its figures and its bounds: » the evidence page · the record itself, kept as a doc.html
every section carries its hasha witness is the SHA-256 of one section's exact bytes. recomputing a SHA-256 takes a stock library and nothing else. which bytes to hash is the one thing the format has to tell you.
$ python tools/verify.py examples/memory.doc.html sections: 12 verified 12/12 sections (mismatches: 0, missing: 0)
no key · no login
nothing to expire
two reference readers ship with the format, built differently on purpose — a document must convince both: [ verify.py ] · [ verify.mjs ] — and the wiki has its own pair, checking the shelf: root, leaves, pins: [ verify_wiki.py ] · [ verify_wiki.mjs ]
write your own — three ways innode tools/build-doc.mjs sections.json out.doc.html
Read SPEC.md and build a conformant doc.html for the content in [your source]. Implement every MUST, pass the Validation Matrix (§10), and satisfy the Definition of Done (§11).