Your agent's memory plugin may be quietly steering it with stale facts

The product called "memory" is a lottery

In an October 3, 2026 post, developer Liao argues that every agent memory plugin on the market works the same way: it mines your session transcripts, generates snippets, stores them in a vector database, and on each prompt attaches the five most similar ones. If the agent is still confused, it gets a tool to search for more. His claim is that this architecture cannot make an agent understand your project, however many features get bolted on top. These are his claims about the market, not independently verified findings.

How it works, and where he says it breaks

RAG stands for retrieval-augmented generation: before the model answers, a system fetches text that looks relevant and pastes it into the prompt. In a vector database, each snippet is converted into a list of numbers (an embedding) that represents its meaning, and "similar" means the numbers sit close together. That measures resemblance, not truth.

From that, the post lists several failure modes:

• Similarity is not correctness. Search ranks closeness in embedding space. It cannot tell you which snippet is current or what is missing.
• Snippets lose context. Motivations, lessons and environment do not fit in a fragment.
• The past is treated as truth. The codebase changes daily, so how accurate are 500 snippets about authentication?
• Agents cannot search for what they do not know exists.
• The store is unauditable. With 10,000 embeddings in SQLite, which ones are stale, never retrieved, or wrong and quietly shaping behavior?

The alternative: write it down

The post's proposal is document-based memory. People do not rewatch a three-year-old meeting to recall a constraint; they consult written records. Likewise, the agent gets a structured Markdown workspace holding instructions, specs, decisions, research and indexes. The loop changes from prompt, build, forget to prompt, consult, build, update. Before working, the agent reads the relevant documents. Afterward, it revises stale ones and adds missing ones while the full picture is still in context.

The author says he began with an internal/ folder and a few rudimentary instructions over a year ago, which grew into a free, open-source plugin called Operator Memory. He says it uses no vector databases, embeddings, summarizers or background daemons, only Markdown files you can read, edit, commit and share. He also notes that AGENTS.md files already work, but are often a project's only documentation.

The stakes

The exposed group is anyone running an agent on a fast-changing codebase through a memory plugin. If the author is right, those teams have an opaque store they cannot inspect, feeding their agent facts that may have expired, with no obvious way to find out. The post also describes people shipping features without reading the code, which is when documentation matters most and gets written least.

Those who gain are teams that can read what their agent knows. A folder of Markdown can be reviewed in a pull request, diffed, corrected by a human, and handed to a new hire. That is a governance property as much as a technical one. Note the source is one practitioner describing his own tool and experience; it offers no benchmarks comparing the approaches.

Questions You Should Be Asking

• Can anyone on your team list everything your agent's memory currently contains, and which entries are stale or wrong?
• When the codebase changes, what process invalidates the memories that described the old version?
• If your agent made a bad decision last week, can you trace it to the specific memory or document that influenced it?
• Who reviews what the agent writes into its own documents, and what stops it from recording a mistaken conclusion as settled fact?
• What evidence does the vendor or author offer, beyond personal use, that their approach outperforms the alternative on your kind of project?

What To Watch Next

The signal is evidence, not enthusiasm: head-to-head comparisons of snippet-retrieval plugins and document-based workspaces on the same long-running projects, and whether memory vendors start offering ways to inspect, edit and expire what they store. If auditability becomes a standard feature request, the architecture argument is being won.