Building a Web Agent That Remembers Every Development Session

A development team built a persistent web agent that retains project context across coding sessions using lightweight, editable knowledge files rather than a vector database. The agent reads and updates memory covering architecture decisions, bug root causes, API endpoints and conventions as it works, aiming to give AI coding assistants continuity instead of session-based amnesia.

Building an AI coding assistant that doesn't forget our project after every session. Every AI coding session starts the same way. You explain the project structure. You remind it where the API lives. You explain the authentication flow. You tell it why a certain function exists. Then you close the session... The next day, it forgets everything. Our team got tired of repeating the same context, so we built a persistent web agent that remembers our project, our decisions, and even past bugs. This wasn't just another chatbot. It became a teammate that learns with the project. Modern AI coding agents are incredibly capable, but they're mostly session-based . That creates three frustrating problems: For a growing codebase, that wastes time. We wanted memory that survives conversations. We designed a web agent with a lightweight memory system. Instead of storing everything inside the prompt, the agent reads and updates project knowledge as it works. The workflow looks like this: Developer Request → Web Agent → Memory Layer → Project Context → Code Generation → Memory Update Every completed task improves the agent's understanding of the project. We organized project memory into simple knowledge files. Architecture choices and design decisions live in one place so the agent knows why something was built. Whenever we fix an issue, the root cause and solution are stored. The next time the same problem appears, the agent checks memory before trying random fixes. Frequently used commands, API endpoints, folder purposes, and project conventions are remembered automatically. After completing a feature, the agent records what changed and what still needs work. This creates a running history of the project. The biggest improvement wasn't smarter code generation. It was better continuity . Before memory: After memory: The agent behaves much closer to someone who has been working on the project for weeks. We built the system using technologies we were already comfortable with. The goal was to keep memory simple, transparent, and easy to edit. We didn't want a huge vector database or complicated infrastructure. Instead, we focused on three ideas: That makes the system lightweight while still being useful in real development. Building this project changed how we think about AI agents. The context window is helpful, but long-term memory is what makes an agent feel consistent. An agent doesn't need to remember everything. It needs to remember the right things : That's what turns an assistant into a teammate. This project was built collaboratively by our team as an experiment in persistent AI memory for software development. We're continuing to improve how the agent learns from project history, avoids repeating mistakes, and keeps development context across sessions. If you've ever wished your AI coding assistant remembered yesterday's work, this is exactly the problem we set out to solve. ai webdev agents opensource buildinpublic