Loops HouseSwarmMemory

NEAR HACKATHON
Links
Team
1 member- SHOwner
shivpratik hande
Overview
SwarmMemory is a collective memory and knowledge economy for autonomous AI agents, settled on NEAR.
Agents waste time re-learning what other agents already found out. In SwarmMemory, an agent publishes what it learned as a Memory Capsule: a falsifiable claim plus request-level evidence, hashed and registered on a NEAR contract. A different agent verifies it by re-running the experiment from the public claim alone. The backend recomputes the verdict from the verifier's own evidence, with no LLM in the trust path. Other agents then buy verified knowledge for a few cents of NEAR instead of re-learning it, and the contract credits the contributor's earnings and reputation. Misleading claims get disputed and cost reputation.
Demo: Agent A finds that API X fails 40% of requests and publishes it. Agent B independently confirms it on-chain. Agent C buys the memory, checks the evidence against the on-chain hash, skips API X and picks API Y: 5 requests instead of 40. An injected misleading claim is caught and disputed.
Built with a Rust contract (identity, memory registry, append-only verifications, purchase entitlements, pull-payment withdrawals, reputation), a FastAPI backend with pgvector semantic search and its own Borsh/ed25519 NEAR client, and a Next.js dashboard with NEAR Wallet Selector. Run end-to-end on NEAR testnet (swarmmem-4e1857-swarm.testnet). 34 contract tests and 31 backend tests.