Why Friday remembers how things felt, not just what happened

Most AI memory stores facts. Friday's also records how each moment felt, because that's what makes the right memory come back at the right time. Here is the research behind it, with our own numbers.

Ask most AI tools what they remember about a customer and you'll get a list: name, last visit, last service. That's a filing cabinet, and a filing cabinet can't tell you which fact matters right now.

A good front desk person remembers differently. They remember that Maria was nervous before her first color. That Dev always runs late and hates being rushed about it. That the owner's Mondays are for her kids, and that booking one would feel like a betrayal, not a scheduling slip.

That second kind of memory is what Friday is built on. Here's why, and how we measured it.

Every memory has to say how it felt

The memory system Friday runs on has one rule the others we studied don't: nothing gets written down without a feeling attached. Each memory carries what happened, plus how it registered: how pleasant or painful it was, how urgent, and a sentence in the writer's own words about what it was like.

We surveyed ten well-known AI memory systems before building ours. None of them required this. They all optimize for one thing: find the stored fact closest to the question. That's the right tool for "what's Maria's phone number". It's the wrong tool for "which customer is about to have a bad day".

What the feeling actually buys

We tested it directly. We took 50 real memories and embedded each one twice, once with its feeling and once without. Then we searched for them two ways:

• By topic, using the memory's title.
• By feeling, describing the emotional content in completely different words, so a match couldn't come from reused wording.

Feelings made topic search a little worse, and made feeling search better every single time, by about seven times as much as they hurt. The two kinds of memory sit on different axes. You don't have to choose one; you have to store both.

That's why Friday can find "the customer who was upset last time" when the question doesn't mention a name, a date, or a service. It's the same way you would.

Why feelings change what she says next

We also ran a small demonstration: the same question ("how are things going?") sent to the same model twice. The first time it had no context. The second time it walked through a handful of related memories, with their feelings, first.

Without the memories it gave the stock answer: "Good, thanks! Just rolling along!" With them, each answer carried the weight of what came before, whether that was warmth, unease, or calm, depending on which memories it walked through.

That was a small test: three trials, one model. We're not claiming it proves anything on its own. But it points at the part that matters for your business: what Friday remembers shapes how she talks to people, not just what facts she can look up.

We measure it from the outside, on purpose

This part isn't glamorous, and it's the part we're proudest of.

A memory system can break silently. Ours did, four times in its first months: once a file was written to the wrong place, once a table quietly emptied, once storage fragmented until search returned nothing. Every time, the system kept answering, and searching by topic still looked fine. What had died was the ability to find things by feeling, and nothing inside the system noticed.

The only thing that caught it, every time, was an outside benchmark running on a schedule. That's the lesson we build on: the memory has to be checked by something that isn't the memory. A system can't grade itself, and we'd rather know.

We found a harder problem too. A feeling written as a label, like "Satisfaction 8/10", passes the rule and says nothing. Hundreds of early memories looked like that, and they made search worse. So we removed that signal and wrote this down as the real bottleneck: memory is only as good as the feelings written into it. It's a quality problem, not a plumbing problem, and it's why Friday is taught to write down what a moment was actually like, in plain words.

What this means for your business

When a regular texts at 9pm, Friday isn't searching a spreadsheet. She's remembering a person: what they came in for, and how it went, and what they'd want to hear.

And when you tell her something that matters to you, she doesn't just store the setting. She remembers what it means.

For technical readers: the numbers above come from our internal research on the memory system Friday runs on: a deployed graph of 8,100+ memories, measured February–March 2026. Topic-mode MRR was 0.39–0.65 on realistic queries, and vector similarity added +0.183 MRR over text-only retrieval. The feeling ablation used N=50, with paraphrased emotional queries to control for lexical overlap.