Memory.Wiki launch drafts

Three pieces of copy ready to publish externally once the founder approves the tone. None of them say "mdfy". All link to the public proof at memory.wiki/mwbench.

The core claim across all of them: one URL works across Claude, OpenAI, and Gemini, verified on content the AIs have never seen during training.

Naming note: the bench project is called MWBench internally, but no one outside knows the name yet. In the drafts below the function is named up front ("cross-AI verification", "open cross-AI eval") and MWBench appears as the project label only when it is the thing being linked to. Lead with what it does, not the codename.

A. Show HN draft

Title (one line, 80 char limit)

Show HN: Memory.Wiki, one URL that works across Claude, ChatGPT, and Gemini

(Alternates if the above feels overcooked:)

  • Show HN: I tested if a knowledge URL actually works across Claude, ChatGPT, and Gemini
  • Show HN: Memory.Wiki, your knowledge as a URL every AI can read

Body

Hi HN,

Memory.Wiki is a personal knowledge hub that publishes every doc, every
curated bundle of docs, and your whole hub as a single URL that any AI
can fetch. No SDK, no plugin, no MCP server. Paste the URL into Claude,
ChatGPT, or Gemini and the AI reads the underlying markdown as context.

I built it because I kept losing answers I'd gotten out of one AI when I
switched to another. Vendor memory layers (ChatGPT memory, Claude
projects, Cursor docs) all live behind walls. None of them help when I
want the architecture decision I made with Claude on Tuesday to be
available to Cursor on Wednesday.

The wedge claim is: "URL paste works across every AI." I wanted that to
not be marketing. So I built an open cross-AI eval to measure it.

The eval runs the same Q&A through three runners (claude-sonnet-4-6,
gpt-5.5, gemini-3.5-flash) against the same hub URL, in two ways:

  - paste mode: the corpus is in the prompt directly
  - browse mode: only the URL plus a fetch_url tool, the AI fetches itself

And against two hub flavours:

  - raymindai: my real public hub (may be in training data)
  - mwbench-zorblax: a fictional synthetic hub I seeded for the test,
    every fact made up (ZorblaxCorp, CipherPlate v3.4.1, Talia Renford),
    impossible to memorize

Results, audited by a quote-evidence judge (every claim must produce a
literal corpus quote):

                  raymindai (familiar)   mwbench-zorblax (unseen)
  paste full        100%                   100%
  paste compact     100%                   100%
  browse            98%                    100%
  tool-use rate     100%                   100%
  adversarial       100%                   (n/a in this run)

The unseen-hub column is the one that matters. AIs can't recall a
hub they never saw during training, so 100% there means the URL
delivery model actually works, not memorisation.

Methodology, harness, and the full nine-round bench log are public at
https://memory.wiki/mwbench. Reproducible locally:

  git clone https://github.com/raymindai/memory-wiki
  cd memory-wiki/eval
  node run-bench.mjs --queries=queries/v1.jsonl

Happy to answer questions about the harness, the corpus format
(concept-clustered digest plus all-docs catalog plus per-doc gist
plus H2 section skeleton), the quote-evidence judge, or where the
wedge breaks. The compact-mode payload is 5 to 9 times cheaper in
input tokens than the full body dump, and it still hits 100% on
unseen content.

Live demo: https://memory.wiki/hub/raymindai
Synthetic unseen hub: https://memory.wiki/hub/mwbench-zorblax
Eval writeup: https://memory.wiki/mwbench
Repo: https://github.com/raymindai/memory-wiki

Anticipated HN questions and prepared replies

  • "Isn't this just publishing markdown to a URL?"Yes, that's the substrate. The product is the structure on top: hub digest, bundle AI graph, per-doc Facts blocks, knowledge graph, all designed so an AI fetching the URL gets navigation, not a wall of text. The eval is what makes the claim auditable.

  • "How is this different from mem0 / Letta?"mem0 and Letta are backend memory stores for AI agents (MCP server, SDK, auto-extraction from conversations). Memory.Wiki is the inverse: human-curated markdown at a public URL that any AI can fetch. Different problem. They optimise single-AI recall; this eval measures cross-AI delivery.

  • "Doesn't 'cross-AI' just mean 'works because the AI can read URLs'?"Right, but it has to actually work. The corpus format has to be navigable by the AI without a custom prompt, the URL has to survive without an SDK, and the answer quality has to hold up against memorisation-free content. The eval measures all three.

  • "Why is raymindai familiar?"raymindai is the founder's public hub, online for months. AI crawlers may have indexed it. mwbench-zorblax is the honest measurement because nothing in it could be memorised.

  • "What's the business model?"Free during beta. Pro tier after with custom domain, password protection, badge removal on shared docs. Document expiry is never a tier feature.

B. Twitter / X thread

Each tweet capped at 280 chars. Numbered for thread order.

1/

Spent a month proving the cross-AI wedge actually works.

Memory.Wiki publishes your knowledge as a URL that any AI can fetch.
Same URL, Claude / ChatGPT / Gemini all read it. No SDK, no plugin.

Today I'm publishing the open cross-AI eval that measures it.

2/

The honest test:

I built a synthetic hub (ZorblaxCorp, CipherPlate v3.4.1, Talia Renford),
every fact fictional, never in any AI training data. Then I asked
Claude / OpenAI / Gemini ten questions about it through the URL.

3/

Result:

  paste full        100% / 100%
  paste compact     100% / 100%
  browse            98%  / 100%
  tool-use rate     100% / 100%

Left column is my real hub (raymindai). Right is the synthetic
unseen one. They match.

4/

The unseen column is the one that matters.

AI can't recall a hub it never saw during training. So 100% there
means the URL delivery model genuinely works, not "memorised the
answer."

100% across three vendors, on truly novel content.

5/

How the judge stays honest:

Every claim in every AI answer has to produce a literal corpus quote.
"Sounds like hallucination" guesswork is disabled. Quote, or the
claim doesn't count.

6/

Compact mode is 5 to 9 times cheaper in input tokens than full body
dump, and still hits 100% on unseen content.

The corpus format: concept digest plus all-docs catalog plus per-doc
gist plus H2 skeleton. Designed so an AI fetching the URL gets
navigation, not a wall of text.

7/

Memory.Wiki is not a memory store. mem0, Letta, and OpenAI Memory
are backend stores for agents to recall.

Memory.Wiki is the URL delivery layer. Your knowledge, in markdown,
at a URL any AI can read. Different problem, different lane.

8/

Full methodology, nine rounds of bench history, and the reproducible
harness:

https://memory.wiki/mwbench

The eval is open. Run it against any hub. The numbers will be the
numbers.

9/

Live demo hub: https://memory.wiki/hub/raymindai
Synthetic unseen hub: https://memory.wiki/hub/mwbench-zorblax
Repo: https://github.com/raymindai/memory-wiki

Free during beta. Capture from any AI, paste your hub URL anywhere.

10/

If you got useful answers out of an AI today and want them tomorrow,
the URL is the answer. The same URL, every AI.

Try it: paste your own hub URL into Claude or ChatGPT after capturing
a few docs.

(end)

C. Short copy variants

Three-line summary (slack, email, OG description)

Memory.Wiki is the URL delivery layer for your AI knowledge. One URL,
every AI: Claude / ChatGPT / Gemini all read it, 100% accuracy on
truly unseen content. Open cross-AI eval at memory.wiki/mwbench.

One-liner

Memory.Wiki: your knowledge at a URL every AI can read. Cross-AI verified, 100% on unseen content.

LinkedIn post (longer, B2B framing)

After a month of testing, I'm publishing an open cross-AI eval that
measures whether a single URL paste actually works across Claude,
ChatGPT, and Gemini.

The wedge claim for Memory.Wiki is: one URL, every AI. Vendor memory
(ChatGPT memory, Claude projects, Cursor docs) all stays inside one
tool. Memory.Wiki lives at a public URL you control, and every AI
reads it the same way.

To make that claim auditable, I built a synthetic hub of fictional
content (ZorblaxCorp, CipherPlate, Talia Renford) that no AI could
have memorized. Then ran the same ten questions through three
runners.

Result: 100% accuracy across Claude, OpenAI, and Gemini, on
content none of them have ever seen. The cross-AI URL delivery
model genuinely works.

Methodology and reproducible harness: https://memory.wiki/mwbench
Live demo: https://memory.wiki

Beta is free. The audit trail is open. If you ship knowledge into
AI workflows and want it to survive vendor switching, this is the
shape.

Publish checklist (founder action)

  • [ ] Read drafts above, edit voice to founder's preferred tone

  • [ ] Pick Show HN title from the three alternates

  • [x] Add screenshot to Show HN (live hub, results table, or the readiness UI)

  • [ ] Post Show HN at a US morning slot (8 to 11 AM ET ideal)

  • [x] Tweet thread same day, hold off two hours after HN to let the comment thread breathe

  • [ ] LinkedIn post for B2B and founder network later in the week

  • [ ] Pin the cross-AI eval tweet to profile until volume settles

Facts

  • Three external launch drafts: Show HN post, 10-tweet thread, LinkedIn post
  • Core claim across all: one URL works across Claude, ChatGPT, and Gemini, 100% accuracy on unseen content
  • All drafts link to the public proof at memory.wiki/mwbench
  • None of the drafts use the legacy "mdfy" branding
  • Drafts lead with "open cross-AI eval" or "cross-AI verification" instead of the project codename, since MWBench is not yet a recognised name
  • Show HN draft includes prepared replies for the four most likely comment threads
  • Tweet thread is 10 numbered posts under the 280-char limit
  • Founder reviews and picks Show HN title from three alternates before publishing