Is memory optional for AI agents, or mandatory?

In 2025, MIT's NANDA initiative published The GenAI Divide: State of AI in Business. Its headline finding travelled fast: about 95% of enterprise generative AI pilots delivered no measurable impact on profit and loss.

The more useful part is the reason. The report's lead author, Aditya Challapally, put it down to how the tools behave at work, not to model quality: generic tools stall in enterprise use because they don't learn from or adapt to workflows. A tool that cannot carry what it learned into tomorrow keeps producing day-one work.

Gartner adds a second warning. In June 2025 it predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and weak risk controls. Gartner does not name memory as the cause. But an agent that re-reads everything on every task costs more, and one that forgets decisions is hard to trust with real work.

The obvious workaround is to paste everything into the prompt. Context windows now hold hundreds of thousands of tokens, so why not give the agent the whole company every time?

Chroma tested exactly this in July 2025. Across 18 models, including GPT-4.1, Claude 4, Gemini 2.5 and Qwen3, performance became less reliable as input length grew, even on deliberately simple tasks. The researchers called it context rot: models do not use a long context evenly.

We saw the same thing on our own team. Our CLAUDE.md, the file Claude Code loads on every turn, had grown to about 23,000 tokens, roughly 30% of everything we sent the model. It cost money on every request and still could not hold the decisions behind our code.

When agents can keep what matters and leave out what does not, the published numbers move in one direction.

• Anthropic. In September 2025 Anthropic released a memory tool and context editing for its agents. In its internal tests, reported by The Decoder, the two together improved agent search performance by 39%, and in a 100-turn web search they cut token use by 84%.
• Snowflake. Adding a plain-text context layer (how tables join, what each row means) to its data agents raised final-answer accuracy by 20% and cut tool calls by about 39%, in an internal test Snowflake describes on its engineering blog.
• Our own test. We asked Claude Code 50 questions about our product, 40 of which needed decisions our team had made in the past. One run had only the repository and its CLAUDE.md. The other could look decisions up from Enkeon when a question needed them.

Memory that is wrong is worse than none. A 2026 study from Databricks, as reported in a review of agent memory research, found that agents which retrieved notebooks from earlier incorrect runs reused those results with even more confidence than before. The agent did not just repeat the mistake. It trusted it more.

So "does it remember?" is the wrong bar. Before trusting an agent's memory with real work, ask:

• When a decision changes, does the old one stop being served?
• When two sources disagree, does someone find out before the agent acts?
• Can you see what it kept, and remove what it should not have?
• Does each person and each AI tool read only what they are allowed to?

How we ran it. Claude Code in headless mode on Opus, read-only tools, 50 questions about our own product: 40 needed past team decisions saved in Enkeon, 10 were answerable from the code. Each question ran once per side and was graded on whether the answer contained the expected facts. The plain run had the repository and its full 23,000-token CLAUDE.md. The run with team context had Enkeon connected and a three-line CLAUDE.md in its place, which is part of why it used fewer tokens.

Limits. It is our own project and our own questions, not a public benchmark. One run per question means some noise. The run with team context lost one code-only question, likely because the long file was gone.