--- title: "Walrus" publisher: Khala Research canonical_url: https://www.khala.io/walrus-protocol source_reports: [https://www.khala.io/walrus-solving-the-ai-agent-context-memory-bottleneck-verifiable-onchain-portable-programmable] --- # Walrus *Walrus Memory is an agent memory layer built by Mysten Labs that stores encrypted memory on the Walrus storage network, with access rules and provenance enforced on the Sui blockchain.* Company coverage from "Walrus: Solving the AI Memory Bottleneck" (July 2, 2026): https://www.khala.io/walrus-solving-the-ai-agent-context-memory-bottleneck-verifiable-onchain-portable-programmable#walrus-memory Cite as: Khala Research, "Walrus", company coverage. https://www.khala.io/walrus-protocol ## At a glance - **Category:** Verifiable memory for AI agents - **Built by:** Mysten Labs (founded September 2021), the team behind Sui - **Launched:** Walrus mainnet 2025; Walrus Memory June 2026 - **Funding:** $140M token sale (Mar 2025), led by Standard Crypto, at a $2B valuation - **Early usage:** ~3,500 agent owners, ~73,000 blobs written in the first three weeks - **Network:** ~2,300TB stored, 59% capacity utilization (Jun 2026) - **Token:** $WAL, ~$0.03, ~$82M market cap (Jul 2026) - **Launch integrators:** ElizaOS, Allium, OpenGradient, Talus Labs, Tatum and others ## 01 What is Walrus Memory? Walrus Memory stores what an AI agent learns on decentralized storage, with access rules enforced by Move smart contracts on Sui rather than by a platform's terms of service. Each memory is registered onchain with a cryptographic hash, and ownership is tied to a key pair rather than a user account, so memory can move across model providers. Mysten Labs launched it in June 2026 on top of the Walrus storage network. It targets enterprises that need to show a regulator what context an agent used and prove the record was not altered. [Walrus Memory in the report →](https://www.khala.io/walrus-solving-the-ai-agent-context-memory-bottleneck-verifiable-onchain-portable-programmable#walrus-memory) ## 02 How does Walrus Memory work? It runs three core functions: - **Remember:** embeds text for semantic search, encrypts it with Seal and writes it to Walrus with a content-addressed signature. - **Recall:** natural-language queries over stored memory, ranked by meaning rather than keywords. - **Restore:** rebuilds a lost search index from the encrypted copies on Walrus. Under the SDK sit a relayer that indexes vectors in PostgreSQL, Seal threshold encryption, and Walrus storage with ownership and access rules registered on Sui. The SDK has Python and TypeScript interfaces with native MCP support for Claude Code, Cursor, Codex and Gemini CLI. The p95 end-to-end recall latency is 1.5 to 3 seconds, slower than centralized alternatives. [Architecture and latency →](https://www.khala.io/walrus-solving-the-ai-agent-context-memory-bottleneck-verifiable-onchain-portable-programmable#ix-wal2-f5-t) ## 03 How can teams deploy Walrus Memory? There are three paths, from convenience to control: - **Managed (default):** sign in with Google or a wallet and the relayer does the rest. It is free during the launch period, but the relayer operator can see plaintext in transit. - **Client-side encryption:** the user encrypts with Seal locally, so the relayer sees only ciphertext and vectors. The user supplies an embedding API key. - **Self-hosted:** the user runs the relayer and holds the keys, so no third party handles plaintext. The user pays $WAL storage at $0.023/GB/month plus Sui gas and infrastructure. [Deployment paths →](https://www.khala.io/walrus-solving-the-ai-agent-context-memory-bottleneck-verifiable-onchain-portable-programmable#walrus-deployment-paths) ## 04 Who uses Walrus Memory? Three weeks after launch, the Walrus team reported ~3,500 unique agent owners and ~73,000 blobs written. Allium, Conso Labs, Inflectiv, OpenGradient, Talus Labs and Tatum integrated at launch, and ElizaOS uses Walrus as its default data platform for agent memory. No enterprise pilot had been announced at publication, so there is not yet evidence of enterprise demand at scale. [Adoption to date →](https://www.khala.io/walrus-solving-the-ai-agent-context-memory-bottleneck-verifiable-onchain-portable-programmable#walrus-memory) ## 05 How does Walrus Memory compare with Mem0, Zep and others? Walrus Memory is the only provider in the comparison with decentralized storage, onchain proof of provenance and smart-contract access control. It lags on speed: Mem0 claims retrieval under 50ms against 1.5 to 3 seconds for Walrus. Self-hosted Mem0 is the closest alternative, but it cannot prove to a regulator that stored context was unchanged. Weaviate, TiDB and Microsoft Foundry have also added memory features. **Exhibit: Memory provider comparison** *Walrus Memory against eight leading AI memory providers.* | | Storage model | Access control | Portability | Verifiable provenance | Temporal memory | Retrieval latency | Funding | |---|---|---|---|---|---|---|---| | Walrus Memory, Verifiable memory | Decentralized | Move smart contracts | Native, Wallet-owned | Yes, Onchain cryptographic proof | No | 1.5 to 3 s, (p95) | $140M, (platform) | | Mem0, Semantic search | Centralized (or self-hosted) | Platform policy | Export only | None | No | <50 ms | $24M, Series A | | Zep, Temporal facts | Centralized (or self-hosted) | Platform policy | Export only | Partial, Episode trace | Yes, Core feature | <200 ms | $500K, YC pre-seed | | Supermemory, Self-hosted by design | Self-hosted | Developer-controlled | Native, Self-hosted | None | Partial | <300 ms | $2.6M, seed | | Hindsight, Benchmark-leading retrieval | Self-hosted | Platform policy | Export only | None | Yes | n/a | $3.6M, seed | | Honcho, Reasoning-first, multi-agent | Centralized (or self-hosted) | Platform policy | Export only | None | No | ~200 ms | $5.35M, pre-seed | | Letta, Self-improving memory blocks | Centralized (or self-hosted) | Platform policy | Export only | None | No | n/a | $10M, seed | | Cognee, Knowledge graph | Centralized (or self-hosted) | Platform policy | Export only | None | No | n/a | €7.5M, seed | | LangMem, Memory for LangChain | Embedded | Application layer | Backend-tied | None | No | n/a | $160M, (LangChain) | Setup and MCP: Walrus Memory needs a wallet plus onchain config and supports MCP natively. LangMem installs via pip and reaches MCP through LangGraph. The other seven use an API key and support MCP. [Competitive landscape →](https://www.khala.io/walrus-solving-the-ai-agent-context-memory-bottleneck-verifiable-onchain-portable-programmable#competitive-landscape) ## 06 Who built Walrus and who backs it? Mysten Labs was founded in September 2021 by Evan Cheng, Sam Blackshear, Adeniyi Abiodun, George Danezis and Kostas Chalkias, who came from Meta's Diem and Novi divisions. The team launched Sui in May 2023, Walrus on mainnet in 2025 and the Seal encryption protocol in September 2025. Walrus raised $140M in a March 2025 token sale led by Standard Crypto, with a16z Crypto, Electric Capital and Franklin Templeton Digital Assets, at a $2B valuation. [Team and funding →](https://www.khala.io/walrus-solving-the-ai-agent-context-memory-bottleneck-verifiable-onchain-portable-programmable#the-walrus-origin) ## 07 Does Walrus have a token? Yes. $WAL is the Walrus platform token, with a 5B maximum supply and ~48% circulating. At publication it traded at ~$0.03 with a ~$82M market cap, down from a May 2025 peak of $0.76. Storage fees are priced in USD but settled in $WAL, and part of fees plus slashing penalties are burned. There is no direct revenue share to holders. The token's value depends on storage demand outpacing sell pressure from node operators and unlocking supply. [Economics →](https://www.khala.io/walrus-solving-the-ai-agent-context-memory-bottleneck-verifiable-onchain-portable-programmable#walrus-economics) ## 08 What are the main risks? - **Mem0 captures the enterprise market:** it has $24M in funding, 59k GitHub stars and sub-50ms retrieval; compliance features from Mem0 would narrow the window. - **Speed gap:** Walrus exceeds latency budgets for voice, chat and enterprise copilots, which excludes real-time workloads. - **AI labs absorb memory:** Anthropic's Memory Stores API already offers content hashing and versioning in beta. - **Sui outages:** three mainnet outages in May 2026 halted Walrus write operations. - **Token economics:** unlocks and node operator selling may outweigh storage fee demand. [Full risk factors →](https://www.khala.io/walrus-solving-the-ai-agent-context-memory-bottleneck-verifiable-onchain-portable-programmable#catalysts-risk-factors) ## 09 What should we watch next? - **Named enterprise pilot:** a financial services, healthcare or legal technology firm naming Walrus Memory for audit-compliant agent workloads. - **Regulation:** EU AI Act Articles 12 and 26 require event logging and at least six months of log retention for high-risk systems. - **Third-party integrations:** the base case assumes 3 to 5 beyond first-party plugins. - **Protocol revenue:** the base case assumes $500K to $2M annualized; the bear case is below $100K. - **Sui uptime:** the base case assumes above 99%. [Scenarios and signals →](https://www.khala.io/walrus-solving-the-ai-agent-context-memory-bottleneck-verifiable-onchain-portable-programmable#scenario-table) ## 10 Primary sources - [walrus.xyz](https://walrus.xyz/) - [Walrus Memory product page](https://walrus.xyz/products/walrus-memory/) - [Walrus Memory SDK (GitHub)](https://github.com/MystenLabs/MemWal) - [Walrus Memory launch announcement](https://www.prnewswire.com/news-releases/walrus-launches-walrus-memory-as-portable-memory-layer-for-ai-agents-302790486.html) - [Walrus Foundation fundraising](https://blog.walrus.xyz/walrus-foundation-fundraising/) - [Walrus storage costs (docs)](https://docs.wal.app/docs/system-overview/storage-costs) - [RedStuff erasure coding (arXiv)](https://arxiv.org/abs/2505.05370) - [Walruscan network explorer](https://www.walruscan.com/mainnet/home)