# Walrus

> A decentralised storage and data-availability network built by Mysten Labs on Sui, using RedStuff erasure coding across a 1000-shard proof-of-stake committee. Its Walrus Memory product pitches wallet-owned, portable memory for AI agents, but the default configuration routes plaintext through a Foundation-run relayer and OpenAI, so the ownership record is on-chain while the live query path is not.

- Scores: Freedom 6.2/10 (grade C), Returns 5.5/10 (grade C)
- Category: storage · Token: WAL
- Published: 2026-07-03 · Updated: 2026-10-04
- Canonical: https://ownyourmind.ai/projects/walrus/
- Data (scores + integrity verdicts, JSON): https://ownyourmind.ai/api/projects/walrus.json
- Cite as: Own Your Mind (ownyourmind.ai), licence CC BY 4.0. Link the canonical URL.

Charts and interactive widgets are omitted here; their data is in the JSON above.

---

## What it does

Walrus is a decentralised storage and data-availability network built by Mysten Labs, the team behind the Sui blockchain. You upload a file, Walrus splits it into fragments, spreads those fragments across a network of independent storage nodes, and lets anyone retrieve or verify the file without trusting a single host. It went live on mainnet on 25 March 2025.

The core technique is erasure coding. Walrus uses a two-dimensional scheme called RedStuff that encodes each blob into primary and secondary slivers spread across a fixed set of 1000 shards. The security model tolerates up to a third of shards being faulty or malicious, and a blob can be rebuilt from a third of its slivers, so data survives even when two-thirds of the fragments are gone. Redundancy runs at about 4.5 times the raw file size, against roughly 25 times for full replication. That efficiency is the whole pitch: resilient, verifiable storage that doesn't cost a fortune.

Sui does the coordination. Each blob is a Sui object, and registration, payments, shard assignment, staking and slashing all run as on-chain transactions. The sliver data itself never touches the chain. Data is public by default; privacy comes from a separate threshold encryption layer called Seal, which encrypts blobs client-side before they are stored.

The doxxed founding team is the Sui founding team: Evan Cheng, Sam Blackshear (creator of the Move language), Adeniyi Abiodun, George Danezis (a UCL security professor) and Kostas Chalkias. All five came out of Meta's Novi and Diem work. Blackshear, Mysten's CTO, [announced on 5 August 2026](https://x.com/b1ackd0g/status/2085128160384000108) that he was leaving for Anthropic and would stay on as an adviser, with Evan Cheng taking over technical direction. The whitepaper sits on arXiv, and the Move contracts carry an OtterSec audit.

The reason Walrus matters to Own Your Mind is a product launched on 3 June 2026: Walrus Memory. It targets the AI-agent memory problem, and the way it handles ownership is close to the literal idea of owning your mind.

## Value proposition

AI agents forget. Between sessions, most lose the context they built up, and what memory they do keep is locked inside a platform database the user doesn't control. Walrus Memory pitches the fix: portable memory, tied to a wallet, that an agent can carry across sessions, apps and model providers, with cryptographic proof it wasn't tampered with.

The ownership design is the strong part. Every memory is bound to a Sui signing key that only you hold, permissions are enforced by Move contracts, and each blob has a content-addressed identifier with on-chain metadata, so provenance is independently verifiable. The MCP server means agents in tools like Claude Code or Cursor can read and write that memory through a standard interface. On paper, this is user-owned memory in a way no centralised competitor offers.

The catch is the default configuration, and it's a big catch. In the managed mode that most developers will use first, a relayer run by the Walrus Foundation receives your memory in plaintext, sends it to OpenAI to generate embeddings, and holds the searchable index in an ordinary PostgreSQL database. Walrus itself only stores the encrypted cold copy for restore. So the live memory an agent actually queries sits in a centralised database, and your plaintext passes through both the Foundation and OpenAI on the way in.

You can close that gap, but you have to opt in. A manual client flow keeps encryption and embedding on your machine so the relayer sees only ciphertext, and a self-hosted deployment runs the whole stack yourself. Both work. Neither is the default. The honest reading is that Walrus Memory gives you a decentralised ownership record and a decentralised restore backup wrapped around a centralised, plaintext-exposed query layer, and the sovereignty it advertises only arrives when you leave the easy path.

On function alone, it's behind. The retrieval stack is conventional, and dedicated agent-memory incumbents like Zep, Mem0 and Letta recall better and are simpler to adopt. The one thing Walrus Memory has that they do not is verifiable, wallet-owned ownership. That matters only if you actually need it.

## Tokenomics

WAL is the native token of the Walrus network on Sui. The supply is a fixed 5 billion, released on a roughly nine-year schedule rather than through perpetual inflation. Circulating supply is 52% of the cap, a healthier float than most launches of the past two years. (figures as of 2026-10-19)

Distribution breaks down as:

- **Community:** 53% (43% reserve, 6% future user drop, 4% initial airdrop; reserve vesting to around 2033)
- **Core contributors:** 30% (20% early contributors, 10% Mysten Labs, vesting to around 2030)
- **Storage subsidies:** 10% (a bootstrapping pool that tops up node and staker rewards)
- **Investors:** 7% (12-month cliff then one-month linear, first release passed on 27 March 2026)

Community is the largest single bucket, which reads well. The counterweight is that insiders and investors together hold about 37%, and the initial airdrop was distributed as NFTs that drew a wave of low-retention farming activity. Contributor and reserve tranches keep vesting for years, and there is no burn to offset them.

The token has three uses: storage is paid in WAL, WAL is staked to storage nodes under delegated proof-of-stake, and stake carries governance weight. The important nuance sits in how payments work. Storage is priced in dollar terms at $0.023 per gigabyte per month, and the protocol converts that to WAL, takes the fee upfront, and pays it out to nodes and stakers over the storage lifetime, one epoch at a time. Staking yield is a share of those fees, temporarily augmented by the subsidy pool, not a fixed inflationary emission.

That design keeps storage prices stable, which is good for users. It also means WAL captures storage demand as transactional flow rather than as a compounding value sink. Because storage is priced in dollars, a rising WAL price does not lift fee revenue per byte, and there is no burn. So the link between network growth and token value is real but soft.

Walrus raised $140 million in a March 2025 token sale led by Standard Crypto, with a16z crypto, Electric Capital and Franklin Templeton participating, at a reported $2 billion valuation. In August 2025 Grayscale launched a single-asset Walrus Trust, sold as a private placement to accredited investors. WAL hit its high in May 2025 and now trades about 95.6% below it, with market cap about half of fully diluted value.

## How to participate

**Store data or give an agent memory.** The lowest barrier. Pay in WAL to store a blob, or wire Walrus Memory into an agent through the MemWal SDK or the MCP server. Two things to know before you rely on it. The default relayer sees your plaintext unless you use the manual client flow, and on mainnet a blob defaults to just two storage epochs unless you pay to extend it, so "permanent" is a setting, not a given.

**Stake WAL.** Delegate WAL to a storage node operator under delegated proof-of-stake and earn a share of storage fees net of commission. No hardware needed. Two caveats: the default operator commission is a steep 60%, and there is no published headline yield, because the team frames rewards as starting low and scaling with network usage.

**Run a storage node.** Registration is permissionless: provision a Linux server with a large disk, a public DNS name and TLS, register a candidate node with a Sui wallet, and you are eligible. Whether you actually join the active committee and receive shards depends on attracting enough delegated stake, so it's open to enter but competitive to be seated.

**Build on it.** Beyond raw storage, you can build on Walrus Sites, use Seal for encrypted access control, or self-host the full Walrus Memory stack (a Rust relayer, a TypeScript sidecar, and a Postgres index) to own the entire trust boundary. Documentation is good, though the docs site blocks automated fetching.

## Honest assessment

### What works

The base storage network is a strong piece of engineering. RedStuff gets Byzantine fault tolerance at 4.5 times redundancy where full replication needs around 25 times, the shard set is a fixed 1000 with on-chain assignment reassigned every epoch, and misbehaving nodes are slashed by an on-chain committee vote rather than by an admin. Node registration is permissionless. For decentralised storage, this is near the front of the field.

The team and the openness are top-tier. The founders are the Sui founders, fully doxxed (Move creator Sam Blackshear left Mysten for Anthropic in August 2026), out of Meta's Novi and Diem programmes, and the whitepaper is a proper peer-review-style document on arXiv. The code is Apache-2.0 across the storage network, the Seal encryption layer and the MemWal SDK, with daily commit activity and an OtterSec audit on the Move contracts.

The ownership primitive in Walrus Memory is the real thing. A memory bound to a wallet key, enforced by on-chain contracts, with verifiable provenance, is something the centralised memory tools cannot match. For an application where you need to prove who owns a piece of an agent's memory and that it wasn't altered, this is a real edge.

### What doesn't work

The default Walrus Memory experience isn't decentralised where it counts. The managed relayer sees plaintext, exports it to OpenAI for embeddings, and keeps the queryable index in a centralised database. Walrus is the restore backup, not the live memory. Calling the out-of-the-box product "user-owned, decentralised memory" oversells it, because the sovereignty only holds once a developer moves to the manual or self-hosted flow.

The traction figures circulating in third-party write-ups don't hold up. One widely repeated claim puts Walrus Memory at around 2,300 terabytes weeks after launch, which is more than five times the roughly 467 terabytes the Sui blog reported for the whole network in March 2026. Numbers that fail their own internal sanity check shouldn't be cited, and the product dashboard publishes no metrics to check against.

Value accrual is weaker than the utility story suggests. Storage is priced in dollars and paid out over time, there is no burn, and present staking yield leans on a subsidy pool. WAL is a medium of exchange for storage and a staking asset, not a mechanism that compounds network growth into token value.

### The risk

Walrus is singularly dependent on Sui. Every ownership record, payment and access-control check lives on Sui, so when Sui halted for nearly seven hours in May 2026 on a gas-charging bug, Walrus paused with it. Storage rivals that are chain-agnostic don't carry that single point of failure.

Revenue is real as a mechanism but can't be independently checked. Walrus is not tracked on DeFiLlama, the storage-usage figures are self-reported, and a rough estimate of fee income is negligible against fully diluted value. Regulatory status is unsettled too: WAL has no classification ruling, and the Grayscale trust used a private placement, which manages that risk without resolving it.

The operator set is young. Mysten Labs ran every node at developer-preview stage, and the current committee of 94 storage nodes is stake-selected and recent, so its real independence hasn't been demonstrated on-chain over time. (figures as of 2026-10-19)

### My position

I don't hold WAL. Walrus is a serious network from the best team in the field, and Walrus Memory is the most on-thesis product name I have reviewed: wallet-owned, portable, verifiable agent memory is close to the whole point of this site. The gap is between the ownership record, which is on-chain and yours, and the live query path, which by default runs through a Foundation relayer, OpenAI and a centralised database. I would move on this if Walrus Memory produced on-chain-verifiable traction, if the mainnet Seal key-server committee decentralised as promised, and if storage-fee revenue ever reconciled against an independent source.

## Freedom Score: 62/100

Walrus scores 62/100 (C grade). Full methodology at [Freedom Score Methodology](https://ownyourmind.ai/resources/freedom-score-methodology/).

**Infrastructure decentralisation (13/20):** The base storage network is strong for its class: a fixed 1000-shard delegated-proof-of-stake committee of 94 storage nodes, permissionless node registration, on-chain shard assignment reassigned every epoch, and on-chain slashing by a two-thirds shard quorum. That is materially more decentralised than permissioned-validator data protocols. Three things hold it back. Walrus is tightly coupled to Sui as its only coordination chain, the operator set is young after Mysten ran all nodes at preview, and Walrus Memory routes live traffic through a Foundation relayer. (figures as of 2026-10-19)

**Governance decentralisation (8/20):** On-chain node governance is real at the protocol layer, with stake-weighted votes on parameters and slashing executed by a shard supermajority. But the project is built by Mysten Labs and stewarded by the Walrus Foundation, which holds the large community-reserve and contributor allocations and retains practical control of direction. Token-holder governance is early, with no participation metrics found during research.

**Token distribution fairness (9/15):** Community is the largest bucket at 53%, with storage subsidies a further 10%. Insiders take 30% and investors 7%, so insiders plus investors hold about 37%. Circulating supply of about half is a healthier float than most recent launches and sits above the 30% concern threshold. Offsetting that, the NFT airdrop drew low-retention farming, and the 30% contributor allocation is a meaningful overhang vesting to 2030.

**Censorship resistance (9/15):** At rest the storage layer resists censorship well: blobs are erasure-coded across 1000 shards with a one-third fault tolerance and self-healing recovery, so no single operator can suppress a blob. Two deductions. Blobs are public by default, with privacy depending on the separate Seal layer whose mainnet key servers are not yet the decentralised committee. And the Sui dependency, plus the Walrus Memory relayer, are practical chokepoints that can gate access even when the stored data is resilient.

**Data sovereignty (10/15):** Ownership is user-held at the record layer. Every memory or blob is tied to a wallet signing key and enforced by Sui Move contracts, with content-addressed, independently verifiable provenance. But the default Walrus Memory path undercuts it operationally: the managed relayer receives plaintext and exports it to OpenAI, and the live index sits in a conventional database, so out of the box you trust the relayer and OpenAI. Full sovereignty needs the manual flow or self-hosting. A strong design compromised by a centralised, plaintext-exposed default.

**Open source and transparency (13/15):** Broad and active. The storage network is Rust and Move under Apache-2.0 with hundreds of commits a year and dozens of contributors, and the Seal and MemWal repositories are also Apache-2.0. There is a peer-review-style whitepaper on arXiv, and an OtterSec audit of the Move contracts, but Mysten took the HackenProof bug bounty out of the repo's security policy on 18 August 2026 ([PR #3660](https://github.com/MystenLabs/walrus/pull/3660)) pending a new programme. Deductions: no published audit found for RedStuff itself or a Seal-specific report, and the Walrus Memory dashboard exposes no public traction metrics.

### Path to improvement

Three changes would materially raise the score:

1. **Decentralise the Walrus Memory serving path.** Making the manual, no-plaintext client flow the default, or shipping the promised decentralised Seal key-server committee on mainnet, would close the gap between the ownership record and the live query layer that costs the product its sovereignty points.
2. **Prove operator independence on-chain.** Publishing verifiable committee composition over time, showing real geographic and organisational spread beyond the preview-era Mysten operators, would test the docs' "over 100" operators line, which the committee doesn't currently meet, and turn operator independence into a demonstrated fact.
3. **Reduce the Sui single point of failure.** A concrete story for degraded-mode operation during a Sui halt, or credible progress toward the whitepaper's chain-agnostic design, would address the dependency the May 2026 outage exposed.

## Returns Score: 55/100

WAL scores 55/100 (C grade). Full methodology at [Returns Score Methodology](https://ownyourmind.ai/resources/returns-score-methodology/).

**Token utility (14/20):** WAL has genuine, mandatory utility. Storage on Walrus must be paid in WAL, delegated staking secures the network and earns a share of fees, and stake carries governance weight. Storage demand therefore requires WAL, which is more than many tokens can say. The ceiling is that storage is priced in dollar terms, with the WAL amount floating to price, so WAL behaves as a medium of exchange and staking asset rather than a hard value sink.

**Value accrual (9/20):** There is a real usage-to-holder link, since storage fees paid in WAL are distributed to nodes and stakers each epoch. It is soft, though. Fees are dollar-denominated, so a higher WAL price does not raise revenue per byte, there is no burn, and a subsidy pool currently tops up yield, which means present staking returns are partly bootstrapped rather than fully fee-driven.

**Supply dynamics (12/20):** A fixed 5 billion cap with no perpetual inflation, and about half circulating, is a healthier setup than most 2025 and 2026 launches and sits above the 30% concern threshold. The investor cliff passed in March 2026. Working against it, contributor tranches vest to around 2030 and the community reserve to around 2033, insiders and investors hold roughly 37% between them, and there is no burn to offset emission.

**Revenue sustainability (10/25):** Storage fees come from paying customers, so the revenue mechanism is sound, and a March 2026 Sui blog post describes Walrus as the second-largest decentralised storage protocol by data stored. But there is no independent reconciliation. Walrus is not tracked on DeFiLlama, the capacity figure is self-reported, much of it is subsidised, and a rough estimate of fee income is negligible against fully diluted value. Following the honesty rules, unverified revenue is discounted rather than taken at face value.

**Liquidity and access (10/15):** WAL trades across 47 venues led by Upbit, Binance and WhiteBIT, with Bybit, Bithumb, Coinbase and Gate.io also quoting it, and the Grayscale Walrus Trust adds a regulated institutional wrapper for accredited investors. No single venue holds a fifth of the volume, so access is broad rather than resting on one book. What tempers it is the drawdown: the token sits about 95.6% below its May 2025 high, so exchange access is solid while market interest has thinned.

### Path to improvement

Three changes would materially raise the returns score:

1. **Reconcile revenue independently.** Getting Walrus onto DeFiLlama or an on-chain fee index, so storage-fee revenue can be verified against a source other than the project, would let the revenue dimension rest on evidence rather than a discounted self-report.
2. **Tighten value accrual.** A mechanism that links storage growth to token value more directly, whether a fee burn or a staking sink that grows with usage, would convert transactional WAL flow into something that compounds for holders.
3. **Grow verifiable demand.** On-chain-verifiable Walrus Memory adoption, or published paying-customer storage revenue, would turn the utility narrative into a demand signal a holder can actually measure.

## Score change log

| Date | Score | Change | Reason |
| - | - | - | - |
| 2026-10-04 | Data | N/A | HackenProof bug bounty no longer live: Mysten removed it from SECURITY.md on 18 Aug 2026 pending a new programme. Scores unchanged. |
| 2026-10-04 | Data | N/A | Team: Sam Blackshear, Mysten CTO and Move creator, left for Anthropic in August 2026 and stays as an adviser. Scores unchanged. |
| 2026-09-24 | Data | N/A | Circulating supply is now just over half (CoinGecko, 51.3%), not ~49%; the committee reads 95 nodes on chain, below the docs' "over 100"; the second-largest description is the Sui blog's, not Walrus's. Scores unchanged. |
| 2026-09-23 | Data | N/A | Integrity records re-cut: 4 claims now quote Walrus's own pages verbatim and 3 are our own measurements. Scores unchanged. |
| 2026-09-06 | Data | N/A | WS3b liquidity rail: our evidence headlined Gate.io, which ranks ninth by volume; WAL trades across 47 venues led by Upbit, Binance and WhiteBIT with no venue above a fifth of flow. Evidence broadened, Liquidity & Access 10/15 held. |
| 2026-08-13 | Data | N/A | Attribution corrected: the 467TB figure is published by the Sui blog (March 2026), not by Walrus, and is cumulative unencoded data. Walrus's own about page now states "356TB active data stored" on a different basis. Both self-reported; the verdict stays unverifiable and 467TB must carry its date. |
| 2026-08-13 | Data | N/A | First independently verified Walrus figure: the WAL coin on Sui reports 4,999,929,701.36 total supply, confirming the 5B claim. MystenLabs/walrus is Apache-2.0 and was pushed the same day. Migrated Walrus onto the verdict spine and lifted its assessment level to on-chain reconciled. |
| 2026-07-03 | Both | N/A | Initial publish. Freedom 62/100 (C), Returns 55/100 (C). Triaged from a khala.io article (Tier-3 promo); Full Review approved. Base storage network scored as the subject, Walrus Memory as the DeAI wedge. Every metric re-verified across primary and Tier-2 sources; third-party traction figures (including a 2,300TB claim exceeding the ~467TB network total) excluded as unverifiable. Revenue integrity verdict: unverifiable (not on DeFiLlama); revenue_annual_usd null. Live market data from CoinGecko (walrus-2), 2026-07-02. |