Unified Agent Memory overview

Unified Agent Memory as a cross-cutting topic — compatibility layers and shared protocols that allow multiple agent products to read from and write to a common memory store. The core problem: each coding agent has its own memory format and location, so a developer switching between Claude Code, Cursor, and Windsurf must maintain separate memory files with duplicated or drifted content. Unified memory layers (memories.sh, agentmemory, memsearch patterns) address this by defining a canonical storage layout that all participating agents can read. The four-type taxonomy (session / semantic / episodic / procedural) is the emerging schema for unified memory, with memories.sh as the reference implementation. Key design tensions: storage format (plain files vs vector DB), update semantics (each agent writes independently vs a central service mediates), and conflict resolution when two agents write different values for the same fact.