I’m posting some thoughts on long term work we’ve been doing at Stak and Sphinx. Various trends are coming to a head and it’s time to share our view on what’s coming next. If you have thoughts or want to work on this with us, reach out over DM or reply below. Can your agent remember? Agents are a key next step for AI to have a real impact on the cost of knowledge work. The likelihood of an agent completing a multistep process is currently very low. To increase those odds, agents need to have record and recall past attempts that worked. The more agents making attempts, the higher likelihood a successful attempt is stored. This ability to store and recall successful strategies across disparate agents is ‘multi-agent memory” (MAM). With MCP, agents already have a standard way to discover and use tools. To avoid repeating past mistakes, there are some key choices to make... Standard or proprietary? MAM is an open standard way for agents to remember which sequence of MCP calls worked. As millions of agents try billions of these sequences, the logical traces they store will determine if an attempt succeeds or fails. Open or closed access? Will this pool of knowledge live inside a corporation, or will we follow the open source model and make it open for the world to use? We are working to make sure it remains open. Protected or exposed? Any open access system exposes itself to spam and garbage inputs. A shared MAM database will require protections against overuse and griefing attacks. Account or incentive based protection? Most public online services like Reddit, Instagram, and others use account based protection. Bad actors can be banned by an administrator. The cost to protect an account based system usually requires ads or subscription fees to pay for the cost of administration. An incentive based system, where the user has to stake some value in order to use the resource, doesn’t have these linear costs. Value staking: credit card, stablecoin, or bitcoin? An incentive system that allows users to stake bitcoin opens up the usage to those who don’t have access to the traditional banking system. It also provides anonymity and fraud prevention. Staking mechanism? The bitcoin lightning network has a system called L402s (ht lightninglabs & buckperley) that uses the web’s native 402 request code to trigger a payment request to the agent. The technology exists today to support a global, protected, shared agentic memory database. The same L402 system can work with MCP tool usage as well. Agent actions will eclipse human activity in the digital world. Agents will need tool discovery via MCP, memory via MAM, and budgets using 402s. All of this works today. If you want to insure an open future, reach out and get involved.