THE MEMORY PROBLEM
Your agents are only as smart as what they remember.
Most agents start every session from zero. Context is scattered, the same user is a stranger on the next channel, and nothing decides what to keep or what to let go. People raise nine objections to fixing this with memory. Here is the fair version of each, and the answer.
Published Updated
A bigger window is not a memory.
The loudest objection is that million-token windows make memory unnecessary. They do not. A bigger window is recall capacity, not memory. It does not fix cost, it does not fix attention, and it resets between sessions.
Memory vs context windows →Bad memory is worse than no memory.
The fairest objection, and we adopt it. A wrong or poisoned memory makes the agent confidently, consistently wrong, at machine speed. That is not an argument against memory. It is the argument for getting it right.
Memory privacy and ownership →01The objections
The nine objections, answered
Each page names one objection, gives it the fair version, and answers it with proof. The same three numbers run through all of them.
THE BIGGER-WINDOW QUESTION
Memory vs context windows
Million-token windows raise the ceiling on how much you can carry, but they do not fix cost, attention, or cross-session continuity. A bigger window is a bigger desk, not a memory.
THE RAG QUESTION
Memory vs vector RAG
RAG retrieves documents. Memory maintains a model of the user and the world over time. We expected vector search to win our own 50,000-document experiment. It did not.
THE SKILLS QUESTION
Memory vs agent skills
A skill file holds facts that stay true and fails on facts that get replaced.
THE BUILD-IT-YOURSELF QUESTION
Build vs buy agent memory
Write it to files, or put it in your database. Both instincts are right that memory must be governed, and both stop short of the operations, entity resolution, decay, contradiction handling, and write-time checks a memory layer adds.
THE BENCHMARK QUESTION
Measuring agent memory
Vendor-graded numbers deserve suspicion, including ours. We publish the methodology, report the losses, and run the pollution, forgetting, and concurrency tests nobody else runs. A number you cannot inspect is a claim, not proof.
THE OWNERSHIP QUESTION
Memory privacy and ownership
Native memory locks your context inside one vendor and leaves your own agents without it. And bad memory is worse than none. The answer to both is memory you own, portable across models and private by design.
THE WRAPPER QUESTION
How agent memory works
The difference between a wrapper and a memory system is the work you cannot see in a demo: entity resolution, contradiction handling, staleness, scoping, and forgetting on purpose. That gap is the difference between memory and a cache with good branding.
Why do AI agents need memory?
Because a model has no state between calls. Everything an agent appears to know within a conversation lives in the context window, and the context window is discarded when the session ends. Without a layer outside the model that decides what to keep and puts it back when relevant, every session starts from zero, regardless of how capable the model is or how large its window has become.
The proof, in numbers
Questions people ask before deciding
Agent memory is state that survives the session: facts, preferences, past interactions and learned patterns, stored outside the model and retrieved into context when they are relevant to the current turn. It is distinct from the context window, which is working memory and is rebuilt from nothing each time.
You can build one, and the question is which parts you want to own. Storage and retrieval are straightforward. Consolidation, contradiction handling, entity resolution across aliases, and deletion that propagates into derived artefacts are where the time goes, and they are the parts that stay hard after the demo works.
Contradictions. The first version stores everything and retrieves by similarity, which works until a user changes their mind and the store now holds both answers with no way to choose. Everything after that is a consequence of not having decided how a newer fact beats an older one.
When sessions are genuinely independent, which is true of one-shot tools, most classification work, and anything where the user is anonymous and does not return. Memory earns its cost when the same person comes back and expects continuity.
The difference between a demo and a product is memory.
Get Started Free