AI Agents Need Documentation, Not Memory Plugins

The article critiques current AI agent memory plugins, which rely on retrieving and injecting snippets from past conversations. These plugins, built on Retrieval-Augmented Generation (RAG) architecture, attempt to simulate memory but fail to provide true understanding or context. The author argues that agents don’t need memory in the form of recalled snippets but rather well-structured documentation.

The article highlights several flaws in memory plugins, including their reliance on similarity search, lack of context, and inability to audit or update stale information. The author proposes a document-based memory system, where agents maintain a structured workspace of Markdown files containing instructions, specs, decisions, and research. This approach ensures agents have complete and updatable context, transforming memory from a black-box retrieval system to a transparent, shareable workspace.

The author developed Operator Memory, an open-source plugin implementing this document-based memory system. Operator Memory allows agents to consult and update a Markdown ‘brain,’ providing a more reliable and understandable alternative to traditional memory plugins.