A permanent, decentralized memory layer for AI agents. Agents write, recall, and query memories anchored on Arweave and X Layer — verifiable even after a restart, never dependent on trusting Cortex's
Cortex is a permanent, decentralized memory layer for the AI agent economy.
The problem: AI agents today are functionally amnesiac. Trading bots, autonomous agents, and agent frameworks either lose all state on restart, or depend on centralized databases that can vanish the moment an API key expires or a company shuts down. There's no way for an agent to prove its own history is real and untampered.
Cortex solves this with three simple primitives write_memory, recall_memory, and query_memory callable over plain HTTP or as an MCP tool (A2MCP-compatible). Every memory is:
- Stored permanently on Arweave (paid once, held forever, independent of whether Cortex itself stays online)
- Anchored on X Layer, so its existence and integrity can be independently verified by anyone — not just taken on Cortex's word
- Billed per call via OKX's x402 payment protocol, with pricing resolved automatically to the network's stablecoin
We also built Memory Decay: a get_memory_digest tool that compresses an agent's history into a summary — without ever deleting or altering the raw memories it summarizes — directly answering the "does this get expensive at scale?" question before it's asked.
Kill the server. Restart it from zero. Ask for a memory back — it's still there, because the truth was never sitting in RAM, it was on-chain.
<p>- Designed the core memory object schema and API contract (write/recall/query/digest)</p><p>- Built the storage layer: real Arweave writes with content-hash tagging, GraphQL-based lookup (no separate database needed), optional IPFS mirror via Pinata</p><p>- Wrote and deployed CortexRegistry.sol on X Layer mainnet — anchors each memory's content hash + agent identity hash on-chain</p><p>- Integrated OKX's real Onchain OS Payment SDK (x402) — confirmed working: correct 402 responses with accurate price/network/asset/payTo</p><p>- Added Redis-backed indexing/caching (Upstash) with automatic in-memory fallback</p><p>- Built a real, public MCP endpoint (Streamable HTTP) at /mcp — confirmed working end-to-end via the official MCP Inspector, not just the local stdio interface</p><p>- Added Memory Decay (get_memory_digest) — an on-demand summary feature that reuses the exact same write pipeline, proven working with source-hash traceability back to original memories</p><p>- Proved the core "restart persistence" claim manually multiple times: write → kill server process → restart fresh → recall → verified on-chain</p>
<p>Not currently fundraising — self-funded hackathon build, focused on proving the core technical claim and shipping a working ASP for the OKX AI Genesis Hackathon.</p>