Customer feedback arrives unsolved. Records, Memories and Workflows are how we solve it.

Most business data arrives solved. A revenue figure means one thing. It's already comparable to last quarter's, it adds up with the others, and nobody argues about how to read it. The work of making it usable happened before it reached you.

Customer feedback arrives unsolved. It's unstructured, contradictory, and nuanced. It comes in several languages. The same thing gets described five different ways by five different people. One customer loves what another can't stand. Someone complains about a feature they asked for last quarter. A five-star review buries a serious complaint in the third sentence.

There is no single right answer sitting inside it waiting to be extracted, because the data genuinely disagrees with itself. That's not a flaw in the collection. It's what feedback is.

The cost of re-solving it

Because nobody solved it upstream, everybody solves it downstream, again and again.

Every project starts from the raw pile. Someone makes judgment calls about what counts, how to group it, which phrasing means the same thing as which other phrasing, how to weigh a complaint against a compliment. Then they produce an answer, and that answer is a snapshot. The analysis you ran in March doesn't help the one you're running in July. The judgment calls get made again, slightly differently, by someone else.

So the same question gets a different answer depending on who asked and when. Not because anyone did poor work, but because the work was thrown away each time.

Solving it once

The alternative is to do that work when the data arrives, keep the result, and keep adding to it.

That's what Customer Memory is. Feedback normalized, so the same thing is described the same way regardless of wording or language. Broken into atomic statements, so each one stands on its own instead of three topics being mashed into one item. Tagged consistently, against your categories, across everything you hold. And augmented with business context such as your CRM, ERP, or knowledge base data.

Contradictions don't disappear – they shouldn't. They become countable. How many people said this, how many said that, which way it's moving, and which segments sit on which side. That's a different object from a pile of quotes, and it's the thing you can actually run a business on.

Memory also compounds, which raw data can't. A weekly report stored as memory is something to compare next week against. An account's history is something to brief from before the next call. The longer it runs, the more it knows, and none of that requires re-reading anything.

"Couldn't a good AI model just do this?"

Fair question, and worth answering properly rather than waving at context limits.

Sometimes, yes. With a few dozen documents and a one-off question, hand them to a capable model and ask. You'll get an excellent answer and you don't need memory for it. We'd rather say that than pretend otherwise.

It stops working as things grow, and it stops working in ways that don't improve with a better model.

Recall degrades on messy data. Retrieval across unstructured, multi-language, inconsistently phrased feedback surfaces the wrong things, not because retrieval is bad, but because what it's searching is inconsistent. Normalizing first is what makes recall reliable.

Classification degrades too. At scale, classification falls back on keyword matching, which misses anything phrased a little differently. Matching on meaning works, but only against normalized text, not raw.

You pay to read everything, every time. Every question means going back over the whole corpus. Not once. Each time anyone asks anything.

The fan-out fix breaks at the reduction step. You can split the reading across sub-agents so all of it genuinely gets read. Their answers then have to be brought back together, and that's where it falls over: too much output to hold at once, too noisy to trust, fidelity lost at every level.

And none of it knows your business. A general-purpose model reads each piece of feedback in isolation. It doesn't know which product line it touches, which segment the customer sits in, what your team is trying to achieve this quarter, or how you talk about any of it internally. Feedback read without that context is just text.

None of this is a gap that closes as models improve. The models already read fine. What decides the quality of the answer is what was done to the data before the question was asked.

What we shipped

Which brings us to this release. Three parts of NEXT have new names — Recordings are now Records, Highlights are now Memories, Automations are now Workflows — and the names are the least of it. Each one moves the product further onto the memory model above.

Records: any feedback, kept whole

NEXT started with calls. Recordings handled audio and video, and highlights were the insights we pulled from them. That worked for calls and less well for everything else: a survey response or review had nowhere to live as raw data, so it came in as a highlight directly. One response in, one highlight out, no matter how many separate things that person actually said.

So a customer who mentioned delivery, packaging and price in one review became a single item covering all three, which is exactly the mashing-together that breaks tagging. A tagging agent works from a set of tags meant to be mutually exclusive and assigns one per item. Given an item covering three topics with three sentiments, there is no right answer. The workaround was a separate tagging agent per tag, which works, but doesn't scale.

Now any raw customer feedback goes in as a record. Audio, video, PDFs, documents, presentations, survey exports. Each type gets its own signal agent that reads it and writes the memories from it, because a research deck isn't the same kind of thing as a support export or a customer call. Your raw data always stays intact, and records scale to millions.

Memories: what's worth remembering

A highlight was a fixed item, and that's where it stopped. A memory is what NEXT, with your instructions, decides is worth remembering from your data, and it keeps building.

From records. Signal agents read each record and write memories from it. A record with nothing much in it produces none. A review touching on delivery, packaging and price produces three. Each one comes out atomic – one topic, one sentiment – and the language is normalized as it's written.

That's the solving-it-once step, made concrete. It's why tagging agents land far more accurately than they did on highlights, why you no longer need one agent per tag, and why keyword tagging works at all: the words being matched are the normalized ones, not whatever a customer happened to type.

From workflows. Any workflow step can store its output as a memory, built on top of the memories from your records. This is what makes memory compound rather than merely accumulate – the difference between replaying every rainy day and simply remembering that an umbrella keeps you dry. A weekly report becomes something NEXT holds onto and compares against next week, without re-running the whole analysis each time.

Workflows: memory put to work

Memory that sits there isn't worth much. Automations are now Workflows, they've moved into the main navigation, and they've gone well past running a simple chat on a trigger.

• Every step produces output you can reference in later steps

• Branch a workflow so it takes different paths

• Store step output as memory and build on it over time

• Run workflows and see each execution with its full output

• Generate and publish beautiful HTML reports straight from a workflow

• Read from Confluence and Google Docs as input, and write back comments or suggest edits into Jira, Confluence, GitHub and more

A workflow can chain 10+ AI chats, compile everything into a page worth reading, and then go do something with it in the tools your team already lives in.

Concretely: a brief on an account or store location, built from all relevant memories, and what changed since the last month, written into Confluence the morning before your customer call or store visit. Nobody opens NEXT to make that happen. Or a segment digest that lands when new feedback arrives. Or backlog items in Jira enriched with what customers actually said about them.

Each of those is only as good as the memory underneath it. A workflow doesn't read your feedback, it reads your memory, which is why it runs in seconds, repeatedly, without re-analyzing anything.

Workflows aren't the only way to call on it. Your customer memory is reachable over MCP, so the AI tools your team already uses can query it directly. Ask Claude what enterprise customers said about onboarding last quarter and it answers from your memory rather than from a pile of transcripts. You can build NEXT workflows from Claude the same way. Same memory underneath, whichever door you come through, and the AI you already know gets a great deal sharper once it's working from what your customers actually said.

This is the part that makes NEXT ambient. The work shows up where the work already happens – in your workflows, in your tools, and now in whatever AI your team already has open.

Chat: still there, and better for it

Chat hasn't gone anywhere. It's still where you go for an ad hoc question, or to try an idea out before committing to it. The natural path is to work something out in chat, then build it as a workflow so it runs without you.

Why it answers differently from a general AI assistant pointed at raw feedback isn't the interface or a better prompt, and it isn't just our proprietary modes that quantify memory either. It's that the question is answered against your customer memory: normalized, tagged, accumulated and resolved, rather than a pile of raw files. Same reason the workflows work.

What stays the same

Everything you had before is still here under the new names. Your recordings are records, your highlights are memories, your automations are workflows, and your tagging agents keep running as they were.

Where this is going

Chat and workflows are two ways to call on your customer memory. They won't be the only ones.

The memory layer is the durable thing — the part that gets better every day you run it, that nobody can build for you, and that other tools can draw on when they need to know what your customers are saying. Solving feedback data once, and then using it everywhere, is the whole idea. This release is us building further into it.