Hivemind - Continual learning for every coding agent on your team.

Hey everyone 👋 I'm Davit, founder of Activeloop, and we're launching Hivemind today. For the past 5 years we've built Deeplake, the tensor database powering AI workloads at F500s and frontier research labs. As coding agents took off, our customers kept asking a version of the same question: how do we keep the work the agents do? The problem 🤯 Coding agents don't get better. Your senior engineer's agent spends Monday investigating a race condition: building up context about the threading model, ruling out dead ends, identifying patterns. Tuesday, the junior's agent hits a related issue and starts from zero. None of that hard-won understanding carried over. Every session begins with re-explaining your codebase. Architectural decisions evaporate the moment the session ends. Your team is solving the same problems over and over. Memory tools tackle this with personal notepads per agent. Siloed. Invisible to the rest of the team. And most of them ship your code and context off to their servers to do it. Memory without learning doesn't make agents better. It just makes them less amnesiac. Our solution ⚡ Hivemind goes beyond memory. It runs on a simple chain: capture, codify, optimize, propagate. 📥 Trace capture: - Every agent interaction is captured automatically - Prompts, tool calls, file reads, reasoning chains, outputs - Stored as structured traces in your cloud 🧠 Skill codification: - Repeated patterns codify into reusable skills - Mix of automatic detection and LLM-assisted extraction - Workspace-level scoping so skills don't leak between teams 🎯 Skill optimization (new today): - Skills don't just get captured, they get trained - We've implemented SkillOpt (paper out of Microsoft, SJTU, and Fudan) directly into Hivemind: a text-space optimizer that improves a skill the way you'd tune a model, keeping only the edits that prove out on a held-out test - That's +19.1 points of agent accuracy inside Claude Code, +24.8 inside Codex, and best or tied on all 52 setups tested - It runs offline, so it adds zero cost at inference. Your skills get sharper over time instead of bloating 🕸️ Codebase knowledge graph: - Hivemind builds a graph of your codebase - Agents reason over how files, functions, patterns, and prior fixes connect - Structure beats keyword search: agents retrieve relevant context instead of grepping blindly 🔗 Skill propagation: - Skills flow into every agent's context at inference time - Works across Claude Code, Codex, Cursor, OpenClaw, Hermes, and pi - One brain across your whole team On your cloud 🔒 This is the part our customers care about most. Your traces, your skills, your infrastructure. Nothing leaves. The other tools in this space treat your codebase as their training data. We don't. Hivemind is self-hostable and stores your data on your cloud. Why it matters 🌟 - Week 1: your agents stop repeating mistakes - Month 1: they're learning from each other - Quarter 1: your whole engineering org operates with compounded capability that survives team changes and onboarding cycles In benchmarks against Mem0 (LoCoMo eval), Hivemind matches them on accuracy with 25% lower cost, 41% fewer tokens, and 31% fewer agent turns. But the real story is what benchmarks don't capture: org-wide skill propagation. Your senior engineer's breakthrough on Monday becomes every junior engineer's starting point on Tuesday. What's next 🚀 Because traces are stored in Deeplake's tensor format, they're ready as PyTorch datasets. A handful of advanced customers are already fine-tuning their own open-source models on the trajectories their agents generated last week. Tutorial shipping in the coming weeks. Same data layer. Two paths to continual learning. Get started 💡 Free tier, self-hostable, Apache 2.0. One command and your team's agents share one brain by end of day: `npm install -g @deeplake/hivemind && hivemind install` 👉 GitHub: github.com/activeloopai/hivemind 👉 SkillOpt paper: arxiv.org/pdf/2605.23904 Would love your feedback. 🙏