MemWal: persistent agent memory
Walrus launched MemWal — a persistent memory layer for AI agents — on devnet and Sui ecosystem builders are already wiring it into agent and trading stacks (examples include ai traders like @0xbeepit). The move targets long‑running agent state rather than ephemeral prompts. (x.com) (x.com)
Walrus published MemWal to devnet as a beta, describing it as a persistent, verifiable memory layer for AI agents in a public product post and documentation. (blog.walrus.xyz) MemWal’s reference architecture routes an AI agent through a MemWal SDK into a backend relayer that writes encrypted blobs to Walrus for storage and uses Sui for ownership and access control. (blog.walrus.xyz) The relayer flow described in the repo embeds inputs, encrypts content, uploads blobs to Walrus, indexes vector metadata in PostgreSQL, and exposes HNSW-based semantic search for recall. (github.com) Docs and project notes list typed memory spaces — labeled conversation logs, checkpoints, reasoning traces, and knowledge bases — framing MemWal as a long‑running agent state store rather than a prompt cache. (docs.memwal.ai) Sui ecosystem activity already includes autonomous trading projects and agent platforms — for example 0xbeepit’s Agent Trader and Talus’s Nexus framework — which use Walrus or similar agent storage patterns and represent immediate integration surfaces for MemWal. (t.co) The MemWal codebase and SDK are published on GitHub and an npm package is available for early builders, while the team documents self‑hosting relayer options and invites feedback as the beta evolves. (github.com)
