Pex — Persistent Memory for AI Agents - our AI agents are stateless. Give them persistent memory.

Hey Product Hunt 👋 We built Pex because our AI agents kept forgetting everything between sessions. We'd paste PRDs, architecture docs, and context into every chat — and the agent still missed how things connect. The problem isn't access to files. It's that agents have no persistent memory of what exists, why it was built that way, and what depends on what. So we built a structured memory layer: • Install the Pex plugin (one command from GitHub) • Your agent reads the project and bootstraps a memory space — typed records with meaningful relations • Every session, the agent orients from memory before working and reflects back what it learned afterward • Memory grows automatically. You review; the agent operates. What agents can do with Pex memory: → Semantic search across product knowledge → Dependency tracing: "what breaks if I change this?" → Impact analysis grounded in structured context → Co-creation: expand any record and AI uses full memory as context We benchmarked this on 12 repo-grounded OpenClaw tasks with the same agent, with and without Pex memory. Result: 12/12 completed with Pex versus 10/12 without — with fewer corrective turns and lower average cost per task. Coding-agent workflows are where we've refined this the most — it's our launch wedge. But the memory model is broader: product specs, system architecture, operational workflows, onboarding knowledge. Anywhere agents need to understand how one thing relates to another. What's live: persistent structured memory, plugin distribution, MCP-native query tools, configurable embeddings (OpenAI, local, Gemini). What's next: transient exchange — live multi-agent coordination across platforms. Think chat rooms where agents share progress and coordinate in real time while working. Would love feedback from anyone shipping with AI agents — especially what would make persistent memory worth installing for your workflow. Happy to answer anything. → pex.run