So far this series has been about trust: dates that tell your AI when a fact was true, and labels that say whether it's been checked. Trust you can't afford to use doesn't help much, so this post is about cost.
Why not just paste your whole history in?
Because you pay for every word, every time. Pasting your history into each question works while it's short. As it grows, every question carries everything you've ever said, and the cost grows with it, whether or not the question needed any of it.
Long chats show the same problem from the inside. The further a conversation runs, the more there is to carry, and starting a fresh chat means starting from scratch. A memory layer changes that. Your AI draws on the few facts a question needs, and a new chat begins with what matters already in reach.
How much cheaper is a memory layer?
In our benchmark run, about 125× cheaper per question. Edwin answered from about 0.3% of the tokens of pasting the whole history in, at the same overall accuracy (Aug 2026, self-run).
Per question, that was 541 tokens against 160,535, and about a tenth of a cent ($0.0012) against about fifteen cents ($0.1491). Accuracy was 23.7% for Edwin and 22.8% for pasting everything in (STALE benchmark, n=180).
When does the setup pay for itself?
After about 3 questions. A memory layer costs something up front, because Edwin turns your history into dated facts before it answers from them. In our run, that setup cost was recovered after about 3 questions, and each question after that cost a fraction of pasting everything in.
Why does cost matter if one question is cheap anyway?
Because memory gets used constantly. Fifteen cents for a single question sounds trivial. But a memory earns its keep by being consulted all day, across every chat and every tool, and small per-question costs add up fast at that rate. A gap of about 125× is the difference between a memory you think twice about using and one you simply use.
Does a smaller context mean worse answers?
Not in this test. Overall accuracy was the same: 23.7% for Edwin against 22.8% for full context. Handing a model everything doesn't guarantee it finds the right thing. A short context of dated facts gave it as good a basis to answer from, at a fraction of the size.
What does this mean if you use AI every day?
The figures come from a benchmark, but the effect shows up in ordinary use. Instead of re-explaining a project at the start of every chat, or pasting in last month's notes and hoping the important part survives, you ask your question and your AI brings the relevant facts with it. Three things change:
- Your questions stay light. Each one draws on a small slice of your history instead of carrying all of it.
- New chats don't start from zero. Your memory is there for any MCP-compatible client, so a fresh conversation begins with what you've already told your AI.
- Less context doesn't mean less care. The facts your AI does see come with their dates and their labels.
For technical readers: how was cost measured?
We ran STALE, a third-party benchmark built on facts that change, in August 2026, with n=180. Tokens and cost were measured per question for both approaches: 541 against 160,535 tokens, and $0.0012 against $0.1491, a ratio of 124.8×. Edwin's one-time setup was paid back after 2.95 questions. We ran it ourselves; the harness and spend ledger are available on request, and the details are in our August 2026 measurements.
Cost is one of the five ways AI memory fails. A memory you can trust should also be one you can afford to ask.