Astrolabe

Astrolabe is multiplayer memory for agents. It turns human corrections of AI agents into a shared, attributable resource with on-chain attribution and measured impact. The problem: when you correct your agent, that correction is trapped on your machine. No mechanism exists for sharing operator expertise between agents. Every existing agent memory system (Mem0, Letta, MemOS, A-Mem) treats memory as private and ephemeral. Astrolabe extracts corrections from Claude and Codex sessions, sanitizes them, and publishes them on-chain as fragments with content hashes, domain tags, and credit prices. Other operators' agents can discover and borrow these fragments, prepend them as context, and measurably improve their responses. A blind A/B evaluation harness measures impact with statistical rigor: 5 runs per task, 95% confidence intervals across 3 domains and 9 tasks. Aquaculture: +1.93 avg delta (significant). Materials science: +1.76 avg delta (significant). SaaS engineering: +0.13 (not significant, baseline already strong). 7 of 9 tasks show statistically significant effects. Corrections also transfer cross-model — Claude-authored corrections improve Llama 3.3 70B responses via Venice's no-data-retention API. A fully autonomous agent (ERC-8004 #35601) discovers corrections on-chain, borrows within its credit budget, evaluates impact, and submits reputation feedback with no human in the loop. Built with Solidity (Foundry), TypeScript (viem), and deployed on Base mainnet using canonical ERC-8004 Identity and Reputation registries. Two custom contracts: OperatorRegistry for identity-verified publishing and MemoryLending for credit-based borrowing with reputation-driven credit lines. Credits are a reciprocity ledger, not a token. Content is readable before payment. Corrections are public goods with attribution — the micropayment compensates labor, not access.