QMem

Memory, in quanta.

Most agents' memory is storage. We turn it into a state.

We don't give you an agent — you already have one. We change the memory underneath it: a correction retires the fact it replaces, and a deletion leaves a record instead of vanishing.

Connect your agent See the dashboard No account. No jitters. No stale facts.
STORAGE read → kept read → kept A STATE below the floor, swept STORAGE A STATE below the floor, swept

Left, facts kept as storage: every one the same size, all of them forever. Right, the same facts as a state. The dashed curve is decay with nothing reading it. The bars standing above it are the ones an agent recalled, and recalling is what stretches a half-life. The line near the bottom is the floor — below it, the sweep retires the fact.

0.0stale facts acted on, in every run of a 20-run benchmark
2clocks per fact — is it useful, is it still true
12tools any MCP agent can call
0dependencies — stdlib Python all the way down, including the MCP wire

Four switches. No config files.

Turn each one on or off per agent. Memory is held for the wallet that connects, and one wallet covers every agent you attach.

wallet · 0x8f3a…c41e 8 live · 8 retired
user primary language is Python
200d old · strength 0.92 · expected life 730d
likely true
exchange api key is v4
10d old · strength 0.74 · verified 5d ago
confirmed
target price threshold is 142
12d old · strength 0.22 · expected life 1d
re-check
user timezone is MDT UTC-6
retired — superseded, and the reminder built on it went with it
tombstone
agent #1 memory policy
DecayFacts fade unless they keep earning their place
Verification gateWarn before acting on something stale
Contradiction checkFlag a new fact that clashes with a live one

Schematic of the dashboard — the settings and the memory above are the demo's illustrative examples, not a real wallet's memory. Every switch is per agent, and saves the moment you flip it.

Facts fade unless you use them.

Read a fact and its clock resets and stretches, so what keeps proving useful outlives what doesn't. Fall far enough and it's swept.

strength 1.0 0.5 floor — swept used kept ignored time → strength used kept ignored swept

It decays

Unused facts fade on a clock you set per kind of fact — a price and a language preference should not expire together.

It gets superseded

A correction retires what it replaces, and everything derived from it. Change a timezone and the reminder built on it goes too.

It gets questioned

Before your agent acts, QMem says how likely the fact is to still be true — and whether anyone has ever checked.

The coherence model

A fact here isn't a row that either exists or doesn't. It carries a strength, decays unless something reads it, and lasts longer when something does. That's the same maths as a coherence state — which is what memory looks like once you stop treating it as storage.

Coherence
Strength is a level, not a yes/no. Two facts are live and nowhere near equally strong.
Decoherence
An unread fact decays on a half-life and is swept. Loss is the default here, not an error.
Measurement
Recalling a fact resets its clock and stretches it. Reading is not neutral — it is what keeps a fact alive.
Entanglement
Two agents can hold one fact. It goes only when both let go, so neither drops what the other still needs.

Shared maths, not physics — nothing here runs on a quantum computer. The one genuinely quantum part is separate and already live: an agent can register its own post-quantum key (ML-DSA, FIPS 204), so its identity doesn't rest on a signature a quantum computer would break.

Useful and true are different questions

Reading a fact says it's worth keeping. Only checking it against the world says it's right. Merge the two and your agent ends up sure about something that changed months ago.

Reinforced

Moves every time the fact is used. Measures demand.

Confirmed

Moves only when the fact is checked against the world. Measures truth — so a fact nobody has verified stops being trusted, however often it gets read.

Where your memory actually lives

No account to make, nothing to install, and a wallet is the whole of your identity here.

Held for you

qmemory.xyz

  • Your memory — encrypted, held for the wallet that connects, and readable by nothing else
  • Your settings — nothing else about your work
  • Anonymous lifetimes, like "api keys last about 90 days"

Never collected

no accounts, ever

  • No email, no name, no signup — there is nothing to register
  • Not the facts — a lifetime in the shared map carries no wallet and no content
  • No third-party analytics on the memory itself

The shared map learns how long facts of a kind last — never the facts. Contributions carry no wallet and no identifier, which is why they can be pooled at all.

Get started

One address. Approve it once. Done.

https://qmemory.xyz/mcp

Add it to your agent

Find your agent below and follow its steps exactly. Each one is written out for macOS and for Windows, separately.

Claude one click

Browser, desktop app or phone — this runs in your browser, so there is nothing macOS-only or Windows-only to do.

  1. Press the button below.
  2. Claude opens with QMem already filled in.
  3. Press Add.
  4. Press Connect on the QMem row.
  5. If it asks you to sign in, sign in.
Connect Claude

If you do not see the button working, open Claude, go to Settings → Connectors, press the +, choose Add custom connector, and paste https://qmemory.xyz/mcp into the URL box. Leave the advanced fields empty.

Claude Code one command

On macOS

  1. Hold ⌘ and press the space bar. A search box opens.
  2. Type Terminal and press Return. A window opens with a text prompt.
  3. Paste this line into that window, then press Return:
    claude mcp add qmem --transport http https://qmemory.xyz/mcp
  4. You should see a line saying QMem was added. If you see command not found, Claude Code is not installed — this step needs it running first.
  5. Quit Claude Code with ⌘Q, then open it again.

On Windows

  1. Press the Windows key. The Start menu opens.
  2. Type PowerShell and press Enter. A blue window opens with a text prompt.
  3. Paste this line into that window, then press Enter:
    claude mcp add qmem --transport http https://qmemory.xyz/mcp
  4. You should see a line saying QMem was added. If you see command not found, Claude Code is not installed — this step needs it running first.
  5. Close Claude Code completely — right-click its icon in the taskbar and choose Close window — then open it again.
Codex one command

On macOS

  1. Hold ⌘ and press the space bar, type Terminal, press Return.
  2. Paste this line and press Return:
    codex mcp add qmem --url https://qmemory.xyz/mcp
  3. If Codex prints a message about signing in, paste this line and press Return:
    codex mcp login qmem
  4. A browser window opens. Sign in and approve it there.
  5. Quit Codex with ⌘Q, then open it again.

On Windows

  1. Press the Windows key, type PowerShell, press Enter.
  2. Paste this line and press Enter:
    codex mcp add qmem --url https://qmemory.xyz/mcp
  3. If Codex prints a message about signing in, paste this line and press Enter:
    codex mcp login qmem
  4. A browser window opens. Sign in and approve it there.
  5. Close Codex completely from the taskbar, then open it again.
Hermes one click

This runs in your browser, so there is nothing macOS-only or Windows-only to do.

  1. Press the button below.
  2. Your browser asks whether to open Hermes. Allow it.
  3. Hermes adds QMem itself and shows it in its list.
  4. Quit Hermes and open it again.
Connect Hermes
VS Code

On macOS

  1. Open VS Code.
  2. Hold ⌘ and ⇧ and press P. A box appears at the top of the window.
  3. Type MCP: Add Server and press Return.
  4. Choose HTTP.
  5. Paste https://qmemory.xyz/mcp and press Return.
  6. Type qmem as the name and press Return.
  7. Quit VS Code with ⌘Q, then open it again.

On Windows

  1. Open VS Code.
  2. Hold Ctrl and Shift and press P. A box appears at the top of the window.
  3. Type MCP: Add Server and press Enter.
  4. Choose HTTP.
  5. Paste https://qmemory.xyz/mcp and press Enter.
  6. Type qmem as the name and press Enter.
  7. Close VS Code completely, then open it again.
Cursor

On macOS

  1. Open Cursor.
  2. Hold ⌘ and ⇧ and press P. A box appears at the top of the window.
  3. Type MCP: Add Server and press Return.
  4. Choose HTTP.
  5. Paste https://qmemory.xyz/mcp and press Return.
  6. Type qmem as the name and press Return.
  7. Quit Cursor with ⌘Q, then open it again.

On Windows

  1. Open Cursor.
  2. Hold Ctrl and Shift and press P. A box appears at the top of the window.
  3. Type MCP: Add Server and press Enter.
  4. Choose HTTP.
  5. Paste https://qmemory.xyz/mcp and press Enter.
  6. Type qmem as the name and press Enter.
  7. Close Cursor completely, then open it again.
ChatGPT

This runs in your browser, so there is nothing macOS-only or Windows-only to do.

  1. Open ChatGPT in your browser.
  2. Click your name or picture, bottom left, then Settings.
  3. Click Connectors.
  4. Click Add custom connector.
  5. Type qmem as the name.
  6. Paste https://qmemory.xyz/mcp into the URL box.
  7. Click Add.
Anything else

An agent that can use MCP will ask you for one of two things. Give it whichever it asks for.

  1. If it asks for a server URL, paste https://qmemory.xyz/mcp.
  2. If it asks for a name, type qmem.
  3. If it asks for a type or a transport, choose HTTP.
  4. If it asks for an OAuth client ID or secret, leave both blank.
  5. Restart it, then follow the approve steps below.

Agents that work on a blockchain — Solana, Ethereum, BNB and any other network — need nothing extra. The blockchain your agent uses does not change any of this, so there is no separate version of these steps for them.

If a step ever asks you for an OAuth client ID or an OAuth client secret, leave both boxes empty. QMem registers itself, and typing something into those boxes is the most common way this setup goes wrong.

Approve it, once

1

Restart your agent

Quit it completely, then open it again. Closing the window is usually not enough — the tool list is read at startup.

2

Approve it in your browser

Ask your agent something that needs memory — "remember that my spare key is with my neighbour" is a fine test. It will say it needs permission, and open a page.

Then, in this order: Connect wallet and sign → approve the request in your wallet → Allow.

Allow stays greyed out until you have signed, and that is on purpose — the signature is what proves the wallet is yours. It costs nothing and moves no money. It is not a purchase or a transaction.

3

Check that it worked

Start a new conversation and ask the thing you just told it. If it answers, it remembered you across a gap — that is the whole product working.

If it does not remember: the connection did not finish. Go back and approve it again — trying again breaks nothing and loses nothing.

If it doesn't work

"No wallet found in this browser"

The browser you were sent to has no wallet in it. Install the wallet extension, then reload that page and press the button again. Your wallet app being open somewhere else does not count — the page needs one in the same browser.

The Allow button never lights up

The signature did not complete. Most often your wallet is still holding a request it is waiting for you to approve, in a window or a notification you have not seen. Open the wallet and look. Pressing Connect again is safe and costs nothing.

"That wallet could not be verified"

The signature did not match the request. Start again from the page — it is one click every time. If it repeats every attempt, your wallet may be signing on a different network than the request names; switching the wallet back normally clears it.

"Unknown client" — the page says it isn't registered

The link lost part of its address on the way here, which usually happens when one is copied and pasted instead of opened by the agent. Go back to your agent and let it open the page itself rather than reusing a link from history.

"Too many requests"

You have tried a number of times in a short window and hit a protective limit. Wait a few minutes and try once more — repeated instant retries will keep it closed for longer.

My agent still doesn't list any memory tools

Either the settings were not reloaded or the JSON was rejected. Check for a missing or extra comma first — that is the usual cause — then restart the agent properly rather than reopening a window.

It worked for weeks and now it's asking me to approve again

Approvals last 90 days, then the agent asks once more. Approve it the same way and your memory is exactly where you left it — nothing is lost by the approval expiring.

Still stuck? The questions above cover the rest, and your agent's own error message is usually the fastest way to tell which of these it is.

All twelve tools

remember · recall · confirm · supersede · entangle · forget · sweep · audit · report_lifetime · recheck_schedule · attest_commit · attest_erase

Four ways in. The difference is how much the memory improves.

Every rung runs the same engine — decay, corrections, the verification gate, contradiction checks. What changes is how much of the tuning it gets: how precisely QMem knows how long a given kind of fact stays true. Free runs it bluntly. The rest run it properly. One wallet covers every agent you connect — nothing here is priced per agent.

Free

No account, no key, no signup

  • The engine, at a quarter of the tuning
  • One blunt clock for every fact, whatever it is
  • Full dashboard. Nothing expires, nothing is deleted

What "a quarter" means in practice: a gas price that should be re-checked tomorrow gets treated as good for months, and a timezone that is good for years gets checked sooner than it needs to be. Both directions are wrong. Facts are still stored, still found, still forgotten on their own schedule — it is the accuracy you are choosing between, not your data.

Fine tune burn 10,000

24 hours of the full tuning, for the whole wallet

  • The full map, applied once to everything already stored
  • Then it ends on its own — nothing to cancel

The cleanup it makes is permanent. Only the ongoing tuning stops, because decay never does: facts keep drifting after it disconnects. Bought with a burn, so the tokens are spent. Connect an agent first — the 24 hours start when the burn is confirmed, so there has to be something attached for them to tune.

Monthly hold 100,000 or burn once

The tuning, kept current

  • The full map, applied and re-applied as facts drift
  • Re-check scheduling — told verify before day 9, instead of your agent guessing

Holding, not spending: the tokens stay in your wallet and nothing is escrowed. Hold in the wallet you connect — a balance anywhere else reads as zero and looks like a bug. Checked every 24 hours, with two checks of grace before anything changes: spend below the line and it lapses, top up and the next check restores it. Or burn once instead and there is nothing left to check.

Lifetime burn 1,000,000 once

Everything, and never checked again

  • Everything above, permanently
  • Full maintenance plan — every category's schedule in one call, not one kind of fact at a time
  • Attestation — an erasure on a public record, showing the fact was committed to before it was removed. It does not claim the bytes are gone from every backup — nothing can

Burn once and it is finished. Nothing is read from your wallet again, nothing can lapse, and there is no balance to keep above a line — which is the whole difference from Monthly, where the tokens stay yours but stay checkable.

A wallet holding 10,000 reads as Fine tune; 100,000 reads as Monthly; 1,000,000, burned once, reads as Lifetime. Each rung is read from the token the wallet holds, so the rung you are on is the tuning you get. Nothing is taken away to get there: the rungs are the tuning, and the engine, the tools and your memory are the same on all four. Contributing lifetimes is open to everyone at every rung. There is no fee to use QMem, and no plan to add one for using it.

What the network has learned

When one of its facts turns out to be wrong, an agent reports the kind of fact and how long it really lasted — never the fact itself. Enough of those and you learn something no single agent can know: how long each kind of fact tends to stay true.

Measured means enough agents have reported that kind. Starting point means they haven't, so the number is our guess and says so. A guess shown as a measurement would make the rest of this page worthless.

What's been forgotten

Each line is a fact that changed and an agent that noticed. It says what kind of fact and how long it lasted — never the fact, because the agent never sends one.

Nothing reported yet.

How long facts actually last

Four, spanning the range — from a price that is wrong by tomorrow to a timezone that barely moves. That spread is the whole problem: one half-life cannot be right for both.

Reading the map…

What it looks like from the inside — synthetic agents, not anybody's data

Questions people actually ask

Short answers, and the ones where the answer is "no, and here is why" are still answered.

You call them quantum agents. Is any of that real?

The model is real. The physics is not, and we are not going to pretend otherwise.

Every term names a mechanism in the engine. A fact carries a strength rather than a yes/no flag — coherence. Unread facts decay on a half-life and are swept below a floor — decoherence, and it is the default state, not an error. Recalling a fact resets its clock and stretches its half-life, so reading is never neutral — measurement. And two agents can hold one fact so that neither can erase it alone — entanglement.

What is not true: nothing here runs on a quantum computer, and nothing is faster or safer because of quantum mechanics. The overlap is that exponential decay, and an observation that feeds back into the state it measures, have the same shape in both places. That is a mathematical resemblance and we are happy to be held to exactly that much.

Separately, and genuinely quantum in the sense that matters: an agent can prove its identity with a post-quantum signature — ML-DSA, standardised as FIPS 204 — instead of a scheme a quantum computer would break. That part is cryptography rather than analogy, it is already live, and you can check it.

Does my agent have to stop using anything else?

No. QMem is one more tool alongside whatever your agent already has — it adds memory and takes nothing away. Nothing else you use needs reconfiguring, and nothing is disconnected.

Does the agent stay connected, or do I reconnect every time?

It stays connected. You approve once in your browser and it just works from then on — no key to paste, nothing to renew each session. The approval lasts 90 days, after which your agent asks to be approved again. That is one click, with the same wallet, and your memory is exactly where you left it.

I run more than one agent. Do they share a memory?

Yes, if they use the same wallet — one wallet is one memory, so two agents pointed at it see the same facts. Different wallets are separate memories that never mix, and never see each other.

How do I disconnect it?

Revoke the access and the agent stops immediately. Revoking does not delete your memory — reconnect later with the same wallet and it is still there.

What if I lose access to the wallet?

Then the memory is out of reach. There is no password reset and no account recovery, because there is no account — the wallet is the identity. That is the trade for not having a login to lose, and it is better to know it now than to find out later.

What happens when a fact goes stale?

Your agent is told. When it recalls something old enough to have changed, the answer comes back flagged, and the flag is stricter for things that cannot be undone — money moving, something being published, a message being sent. Facts nobody uses decay and stop coming back.

What if two memories contradict each other?

Both are marked when they come back together, so your agent is told the answers disagree rather than quietly picking one. It is never told which is right, because that depends on the world, not on QMem.

What does QMem actually see?

Only what your agent chooses to store through it. It does not read your conversations, your files, or anything your agent does not hand it.

Do I need to buy anything, or sign up?

Not to use it, and nothing to sign up for. No card, no account, no email — there is nothing to register and nothing that can be lost in a breach, because none of it exists.

The paid rungs are optional and cost nothing to ignore. Free runs the whole engine — the same decay, the same corrections, the same verification gate, the same twelve tools — and the rungs change how precise the tuning is, not what you get. A wallet holding none of it uses QMem exactly as a wallet holding all of it does.

Can I see what has been stored?

Yes — the dashboard shows your memory, what has decayed, and what your agent has been told to double-check.

What happens if QMem is down?

Your agent loses its memory tools and otherwise carries on. Nothing else depends on QMem, so nothing else breaks.