Building a hosted MCP server revealed hard truths about AI agent memory and OAuth

A developer building Sirro, a hosted memory layer for AI coding agents served over MCP, shared four practical lessons learned from running the server in production. A key early mistake was allowlisting OAuth clients by name, which broke access for any agent not explicitly recognized, since clients like Cursor, Claude Code, and Codex each identify themselves differently. The developer found that accepting dynamic client registration and loopback redirect URIs by default is far more robust than a named allowlist in a fast-moving ecosystem. Tool responses were redesigned to return only snippets and metadata in list calls, reserving full asset retrieval for explicit requests, in order to avoid wasting users' limited context window budgets. The project also challenges the current industry norm of using static markdown files like CLAUDE.md for agent memory, arguing that a persistent, queryable memory layer better preserves hard-won coding work across sessions and projects.