Your Agent's Memory Is a Lottery: The Memory-vs-Documentation War of 2026

In October 2026, Meta's AI assistant Muse unveiled a page for every individual in a user's life, refreshed hourly. This documentation-based memory system, marketed as transparency, sparked controversy among privacy researchers who viewed it as creating dossiers on those who did not even install the app. The heart of the debate lies in the nature of agent memory, with engineers now questioning whether agents truly need memory at all.

AI agents operate in a continuous loop of reading a goal, examining the world, taking an action, and reading the result, then repeating. Traditionally, memory was added to agents by scraping old chat logs and transcripts, breaking them into thousands of snippets, embedding each, and storing them in a vector database. On each prompt, the agent fetched the five most similar snippets and incorporated them into the context. If confusion persisted, a search tool would allow the agent to rummage through its own memory.

However, engineer Kevin Liao argued that agents do not require memory, and instead, they need documentation. He criticized the widely used vector similarity measures, stating they only consider the closeness of texts, not their truthfulness. Liao highlighted five fatal flaws stemming from this flawed assumption, including the potential for outdated information to be retrieved instead of the most current version.

Liao proposed a more straightforward solution: a Markdown workspace containing documents and an index that agents would read, update, and review like any other code. This approach eliminates the need for embeddings, vector databases, and background daemons. In fact, a free, BSD-licensed plugin called Operator Memory has already been developed to integrate this approach for various language models.

Recent developments illustrate the shift in the memory debate. Meta's Muse now presents hourly dossiers, a move away from the vector database approach. Another research line focuses on context language models that maintain their working state as writable documents rather than opaque vector stores. The growing backlash against memory plugins signals a market reaction as well.

Ultimately, the future of agent memory may lie in a simple, versioned Markdown workspace that agents can read, update, and review, ensuring knowledge remains accurate and up-to-date.