Context windows now hold around a million tokens, some ten million. But performance drops with every additional token, in every model. Long context files like CLAUDE.md or AGENTS.md, bloated memory files and superfluous rules therefore cost you twice: tokens and intelligence. Context engineering means putting only what matters right now into the context. That is exactly what Hebby is built for.
Millions of tokens. And still forgetful.
The leading models from Anthropic, OpenAI and Google now process around one million tokens at once, some even ten million. That's entire codebases, shelves of books, months of chat history. The obvious idea: just dump everything in.
It doesn't work. A context window is a capacity, not a performance promise. Just because something fits doesn't mean it's used well.
What the research shows
- Every token costs performance. In a study of 18 leading models, performance dropped as input length grew in every single one. Even on the simplest tasks.
- Even at small sizes. In a demanding test, 11 of 13 models fell to half their performance or below at just 32,000 tokens. That's three percent of a million-token window.
- Focus beats volume. When models get only the relevant parts of a conversation history instead of all of it, they answer much better. In every model family tested.
- The middle gets lost. Models use what sits in the middle of a long context worse than what sits at the beginning or end.
The field calls this context rot: the fuller the context gets, the more it "rots". Anthropic speaks of an attention budget that every additional token depletes.
Your agent doesn't get smarter by reading more. It gets smarter by reading the right thing.
Just like humans: working memory is scarce
Our working memory keeps only a few things present at once, and everything beyond that crowds each other out. The context window is the AI's working memory and hits the same limit. The solution is the same: lean working memory, strong long-term memory.
Context files: the silent performance killer
CLAUDE.md, AGENTS.md, GEMINI.md, MEMORY.md: every agent has its context files. They load in every session and sit in the context on every single call. Over time, everything ends up in them.
- They grow. Every line costs tokens on every call.
- They contradict themselves. Old and new rules sit side by side.
- They distract. Only a fraction matters for the current task. The rest is noise and eats attention.
- They need upkeep. Cleaning up becomes a permanent chore.
And because the files eventually stop being enough, the agent re-reads entire logs and notes just to remember. You hardly notice, but it costs a lot.
Too many rules make models worse
Instructions are context too. And here, too, more is not better.
- Rules get lost. With 500 simultaneous instructions, even the best models follow only about two thirds. Early rules are favored, late ones overlooked.
- Strictness backfires. Current models respond so precisely to their instructions that Anthropic explicitly advises dialing back harsh wording like "CRITICAL: You MUST …". It leads to overeagerness instead of better results.
- Rules go stale. Smarter models need fewer constraints. Prescribing what they already do best replaces their judgment with yesterday's rule.
Because models improve at breakneck speed, this problem grows with every update. Yesterday's helpful instruction is today's straitjacket.
Context engineering: the right thing at the right moment
The answer is context engineering. Anthropic sums up the goal: the smallest possible set of tokens with the highest information value. For a memory, that means:
- Knowledge, not rules. What the model can't know: your decisions, projects, experiences. Not how it should do its job.
- Only when needed. Memories arrive when they're needed. Not across the board on every call.
- Condensed. One precise memory beats three paragraphs of notes.
- Up to date. Outdated material disappears before it creates contradictions.
How Hebby keeps context lean
Hebby is built on exactly these principles. Instead of ever-growing context files, Hebby delivers only the few memories that matter for each request, and anticipates what you're about to need. It clears out outdated material by itself.
Hebby doesn't dictate anything to your model that it can do better itself. And when the model improves, Hebby keeps pace automatically. The result: your context files shrink to one or two sentences, the context stays lean and your agent stays focused. Smarter simply by leaving things out. Our measurements show what that means in tokens.
Hebby Memories: the second brain for Claude Code, Codex & co. Learns as you work, thinks ahead, never forgets what matters. Get Early Access →
Frequently asked questions
What is context engineering?
Context engineering means giving an AI model exactly the context the current task needs, and no more. It goes beyond prompt engineering: it covers everything that lands in the context window, including instructions, files, history, tool results and memories.
What is context rot?
Context rot describes how language model performance declines as more tokens fill the context. It applies to every model tested and sets in long before the end of the context window.
How long should CLAUDE.md and AGENTS.md be?
As short as possible. Only what must apply in every single session belongs in them. With a memory like Hebby, one or two sentences are usually enough. Everything else enters the context when needed.
Do more instructions make a model more reliable?
No. The more rules apply at once, the more of them get missed. And with current models, overly strict instructions lead to misbehavior rather than better results.