Nobody holds that. Neither do you.
So every time a question comes up, you paste a few chunks into a chat, re-explain the context, and hope it read the right paragraph.
https://memory.hokmah.dev/mcp
In Claude: Settings → Connectors → Add custom connector, paste it. Ten documents free, no card. ChatGPT, Gemini, Mistral and your own code call the same address.
Click a verdict to see it
“The corridor, 530 km long, remains the province’s main logistics axis.”
The corpus holds 830 km, not 530 km.
Chapter 2 anchor · 2024 survey · settled, supported by three statements. Your 530 km figure comes from a version retired on 5 August.
Your assistant has two faults. The second one is the dangerous one.
You re-paste the same files and re-explain the same context every week. The conversation you had in January is gone.
A figure it retrieved, a figure it inferred and a figure it invented all reach you with identical assurance. Sorting them is your job — every time, on everything.
Not the kind that remembers — those already exist. The kind that stops you when you say the opposite of what you have already established.
You give it the corpus once. It does not store it in bulk: it breaks every document into statements, files them, links them, and keeps track of where each one came from. Then you work normally. It already knows.
Ten documents, a thousand pages, four separate matters that never mix. It reads the tables, the annexes and the page numbers.
Ask a question, get an answer from the corpus — not from whatever you managed to paste in before the context ran out.
Paste your draft. Sentence by sentence: supported, unknown, or contradicted — with the figure it holds instead.
A correction lands. A ruling reverses last year's reading. You drop the new text in.
It does not file it next to the old one. It retires what the new text replaces. The superseded paragraph cannot come back in six months looking like it is still in force — which is exactly what happens in every folder, every drive and every other memory tool you have used.
Standing is recomputed from the shape of your corpus on every write. It is never something a model declared about itself. Each account gets its own database: another account’s data is not hidden from you, it is not there.
Your corpus settles it, and agrees. It shows you which statement, from which page.
Nothing you loaded can settle it. It says so instead of filling the gap with something that reads well.
Your corpus holds the opposite — and hands you the figure it holds instead, with its source.
The second one is why it exists. Every other tool answers. This one is allowed to decline — and a tool that can say “I don’t know” is the only kind whose “yes” means anything.
PDF, Word, PowerPoint, Excel, CSV, plain text. Tables, annexes and page numbers included — so a figure can always be traced back to the page it came from.
A scanned document, it will tell you plainly: “this is an image — run it through OCR first.” It will never pretend to have read an empty file. That refusal is the whole point of the product.
Every statement carries where it came from and what it rests on.
When two things cannot both be true, it does not pick a winner — it shows you both and tells you which one it already held, and since when. A statement's standing is recomputed from the shape of your corpus on every write, and is never something a model declared about itself.
Each account gets its own database — not a filter over a shared table. Another account's data is not hidden from you; it is not there.
Settings → Connectors → Add custom connector https://memory.hokmah.dev/mcp
Not using Claude? ChatGPT, Gemini, Mistral, your own code — there is an address to call and it works the same:
# create an account and get a key curl -X POST https://memory.hokmah.dev/v1/signup \ -d '{"email": "you@example.com"}' # check a sentence against everything you hold curl https://memory.hokmah.dev/v1/check \ -H "Authorization: Bearer hkm_…" \ -d '{"text": "The corridor is 530 km long."}'
One thing you should know before putting a client file in: automatic decomposition sends the extracted text to an external model. Your file is never stored there, but the text transits. For material under strict confidentiality, have your own assistant read the document and write the statements — nothing leaves your session, and the result in memory is identical.
Same structure, same described function, two lengths, 147 pages apart. Nobody had caught it.
Seventeen minutes. Four cents. The document is public — go and check.
https://memory.hokmah.dev/mcp
Ten documents free, no card. It tells you what it does not know.