Build Day: Agent Harness & Memory [Build Club x Mem0]

[invite only, register for approval] what this is the bottleneck in agentic AI right now isn't the model- it's the harness and the memory. the loop the agent runs in, the tools it can reach, the context you feed it on every turn, and whether any of that state survives past a single session or gets thrown away the second the process exits. SO, lets spend a day on this. the event will run as a few live walkthroughs from people actually building this stuff, then a long sprint where you build your own, then demos. the theme is agent harnesses and memory, but it's an anchor, not a fence. bring what you're working on in agentic space. this event is hosted in collaboration of Build Club x Mem0 how the day runs how the day runs 9:00 – 9:30 → doors, check in. set up and write what you're building today. 9:30 – 10:00 → build day intro. why harness and memory, how the day works, getting the whole room on the same map. 10:00 – 10:30 → research yaps. fast reads from the community on where agent harnesses and memory actually are in 2026. 10:30 – 10:40 → break 10:40 – 11:10 → workshop 1, live in code. Saurabh Mishra- Build Agentic AI App That Your Customer can Interact with via Telegram [Google ADK] This hands-on lab guides you through building a production-ready agentic AI application that connects a Telegram bot to a PostgreSQL database. You will learn to use the Google Agent Development Kit (ADK) and Model Context Protocol (MCP) to enable the AI to perform autonomous reasoning and data operations, all deployed serverlessly on Google Cloud Run. 11:10 – 11:20 → break 11:20 – 11:50 → workshop 2, live in code. Jitendra Gupta- From Memory to Action: Building an Autonomous SRE Agent with Mem0 and OpenClaw Modern AI agents can answer questions, but production systems require agents that can remember past incidents, learn from operational experience, and take meaningful actions. In this hands-on workshop, participants will build an Autonomous SRE Agent using Mem0 for persistent memory and OpenClaw for computer-use automation. Starting from a Docker Compose-based microservices environment, attendees will create an AI agent that investigates incidents, analyzes monitoring dashboards, recalls historical resolutions, and generates actionable remediation recommendations. 11:50 – 12:00 → break 12:00 – 12:40 → build sprint, part 1. the actual point of the day. extend a walkthrough, bring your own project, or take a bounty off the board. 12:40 – 1:20 → lunch. keep building if you're in flow. food and arguments about harness and memory architecture. 1:20 – 2:00 → build sprint, part 2. ship it. 2:00 – 3:00 → demos. open floor, 5 minutes each. present your sprint progress or show a project you've been building. 3:00 – 3:15 → wrap, next steps, and pictures. call for builders 🛠️ [CLOSED] the stage is open for builders to lead these demos and showcase your projects. there are two type of CFP you can apply for: lead a build walkthrough - WORKSHOP take the room through a harness or a memory layer, live, in code. you don't need a polished talk, you need a working build and the ability to explain why it's wired the way it is. secondary slots: - research yap: 10 minutes on the state of agent memory and harnesses in 2026 - demo: show something you built around agents, harnesses, or memory. doesn't have to be from the day. any sharp build counts. submit → https://forms.gle/vpbQ4mKGsHH13Fpy5 [STRICTLY NO SLIDES] (give a walkthrough) bounty board if you don't have your own thing to build, take one of these. each is scoped to ship inside the sprint: harness build the loop. write a minimal harness from scratch: the model calls a tool, you parse the result, then loop until done. walk away with a ~40-line agent you fully understand. tool router. an agent that picks the right tool from eight or so options and shows its reasoning, ending with correct routing plus a log of why it chose each one. context cop. wrap an agent so it summarizes old turns before the context window overflows, with a live token meter, so a long chat never blows the window. harness bench. run one task on two harnesses, ADK and OpenClaw, and score them. end with a scorecard of steps taken, tokens used, and whether each finished. memory second brain bot. feed it facts through the day and it answers later, surviving a restart. build a bot that recalls across sessions, with a view of what it actually stored. goldfish vs elephant. run the same agent twice, memory on versus off, side by side. the result is a demo that makes the case for why memory matters on one screen. recall score. feed facts in session one, ask about them in session two, and count the hits. end with a recall percentage on a small test set you build yourself. decay lab. give memories a timestamp and score, then decay them on retrieval, FadeMem-style. plot what survives and what the agent forgets over time. both telegram concierge. take an ADK Telegram agent and wrap Mem0 read and write around it. the result is a bot that remembers each user across separate chats. sre war-room. extend an OpenClaw and Mem0 SRE agent so it recalls past incidents. send a new alert and watch it surface the matching past fix. many minds. one model with swappable persona and memory files loaded at runtime. switch live between three specialist agents that each know different things. just some inspirations, more on the wall on the day. who's this for builders. learners. bring - laptop, charged, plus charger your stack of choice - an LLM API key (BYO) - a project to build, or an open mind for the bounty board LET'S BUILDD :)