Tokenized Memory: Can AI Memory Become an Asset Class?

This article was prompted by a post from Seb (@sebbsssss) announcing the Portable Memory Protocol (PMP) — “the open standard for AI memory: portable, private, personal.” The framing is sharp: the most valuable standard in AI, Seb argues, is still unclaimed. Context already has one (Anthropic’s MCP). Agents have theirs (Google’s A2A, Coinbase’s x402 for payments). Memory — the layer that, in his words, “compounds with use” — does not. So the race is on to tokenize it.

It’s a good provocation, and a good excuse to do what a tokenomics lab should do with any “tokenize X” pitch: ask what the phrase actually means, and which of its hard problems a token solves versus which it just papers over. This is not a review of PMP — it launches in weeks and there’s nothing to test yet. It’s an analysis of the idea underneath it: tokenized memory as an asset class.

“Tokenizing memory” is three different things

The first job is disambiguation, because the phrase quietly bundles three separate assets with three separate economics:

1. The memory data itself — embeddings, conversation history, learned preferences, the context that makes an assistant feel like your own. Tokenizing this means wrapping a data asset in transferable rights. This is a data-ownership and IP problem.
2. Access and usage rights to that data — who may read or write it, at what price, for how long. Tokenizing this is licensing expressed on-chain. This is a DRM and contracts problem.
3. A network token for the protocol — the implied “memory standard” coin that coordinates and monetizes the marketplace. This is the classic does-this-protocol-need-its-own-token problem.

Most pitches blur the three into one slide. They are not one thing. You can solve (1) and (2) — title and licensing — without ever issuing (3). Keeping them apart is most of the analytical work, so the rest of this piece walks the property list and sorts each item into the bucket where its real difficulty lives.

Walking the property list

Owner. An NFT or registry entry can record title cleanly — this part is solved. But title to what? The data itself can’t live on-chain: it’s too large and, if it’s personal memory, must stay private. So the chain holds a pointer (a hash, a key reference) and the bytes sit off-chain. You own a claim, not the content — the same on-chain-title-versus-off-chain-asset gap that every real-world-asset token has never fully closed.

Provenance. Cryptographic signing and attestation make this genuinely useful, maybe the most useful property on the list. Where did a preference come from — a real interaction, or an injected instruction? For agents acting on your behalf, verifiable provenance of memory is closer to a security primitive than a financial one.

Access rights. Encryption plus key management can gate who decrypts. The catch is fatal and specific: to be useful, memory has to be read by a model — and the moment it is decrypted and ingested, it can be copied. A license can say “read once”; the model’s weights have already absorbed it. This is the same “no trustless link between the token and behavior” problem we flagged for AI agent tokens, one layer deeper: with memory, the thing being sold is information, and information doesn’t stay sold.

Price. Pricing an information good runs straight into Arrow’s information paradox: you can’t judge what a memory is worth until you’ve seen it, and once you’ve seen it you no longer need to buy it. Markets for information goods route around this with reputation, bundling, and subscriptions — not per-unit spot prices. So “every memory unit has a price” is the hard part of the design, not an afterthought the token handles for you. (For the broader question of how AI and crypto try to price a “unit of work” at all, see Units of Work.)

Expiry rules. Time-decay is interesting for two reasons. As a mechanism, it’s a velocity sink — it forces re-acquisition and gives the unit of account something to do. As an idea, it quietly contradicts the headline: memory that expires does not compound with use. You can have decay or compounding; selling both in the same sentence is a tell.

Usage history. An on-chain log of reads enables the one model that genuinely fits: streaming royalties. The owner earns each time the memory is used, rather than once at sale.

The formula is trivial. Its only hard input is Uses — and Uses is honest only if a read can’t become an unmetered internalization. That sends us right back to the access-rights problem: the royalty model is sound exactly to the degree that copying is preventable, which for ingested information is barely.

The three problems a memory token can’t dodge

Strip away the property list and three structural problems remain. They are the reason this is hard, and a serious design has to answer all three on the first page, not the last.

1. Non-rivalry and the copy problem

Memory is information. Information is non-rival — your use of it doesn’t diminish mine — and, once revealed, costlessly copyable. Tokenizing a non-rival good means you are selling licenses, not the asset, and license enforcement over data a model has already consumed is close to impossible. Every durable information market solves this not with better cryptography but with a different value source: provenance, freshness, reputation, exclusivity of relationship. A memory protocol that prices the bytes will leak; one that prices trusted, attested, continuously-updated access has something to sell.

2. Portable versus valuable

The pitch’s two best words are in tension. “Memory compounds with use” is true — but it compounds for whoever aggregates it. That’s a data network effect, and it accrues to the holder, not to a unit floating in a permissionless market. Make memory fully portable and forkable and the moat evaporates; lock it down enough to capture value and it’s no longer portable. The interesting design question is where on that spectrum a protocol sits, and “both ends at once” is not an answer.

3. Personal memory is PII

Tokenizing personal memory collides with data-protection law head-on. The GDPR right to erasure expects data to be deletable; a blockchain pointer is built not to forget. “Expiry rules” gesture at this, but an immutable on-chain record of who accessed what memory when is itself sensitive data, and a liability, not a feature. Any serious version of this lives or dies on a privacy architecture — encryption, zero-knowledge proofs, off-chain storage with on-chain commitments — long before it gets to a token.

So does it need a token?

This is where a tokenomics lab earns its keep. Run the protocol through the standard filter (When You Don’t Need a Token) and the answer is “only under one condition.”

A memory marketplace can settle in stablecoins, record title in NFTs, and gate access with encryption and a reputation registry. None of that needs a native L1-style coin. A native token earns its place only if there is a coordination problem that fiat and NFTs can’t solve — for instance, a permissionless network of independent nodes storing, attesting, and serving memory, which needs crypto-economic incentives and slashing to stay honest. That is a real use case, and it has a name: it’s DePIN. If the memory layer is actually a decentralized storage-and-attestation network, a token is justified by the same logic that justifies one for Filecoin or Arweave. If it’s a centralized service with a marketplace UI, the token is the AI-agent failure mode all over again: narrative wrapped around someone else’s servers.

When tokenized memory makes economic sense

Five questions, in the spirit of the lab’s other checklists. If most answers are “no,” you have a data marketplace with a token bolted on — not a memory asset.

A project that answers all five honestly is rare. Most “tokenize memory” pitches will answer the first as “not really,” the third as “not yet,” and hope the narrative carries the rest — exactly the pattern that turned the AI-agent category into memecoins with extra steps.

Takeaways

1. “Tokenizing memory” is three things — the data, the access rights, and a protocol token. Separating them dissolves most of the hype.
2. The token is the easy part. The hard parts are non-rivalry (information doesn’t stay sold), the valuation paradox (you can’t price what you must reveal to value), and privacy law (personal memory is PII on an immutable ledger).
3. “Portable” and “valuable” pull against each other. Memory compounds for whoever aggregates it; full portability erases the moat.
4. A memory protocol can be a real business and still not need a token. Cash flow first — KYA enterprise contracts — is the right instinct. A native token is justified only if the layer is genuinely a decentralized network (DePIN), not a marketplace with a coin attached.
5. The right question isn’t “is memory the last unclaimed standard.” It’s: what is the unit of value, who can copy it, and does coordinating it actually require a token.