If you’ve ever had to remind Miles of something you told him earlier in the week, this one’s for you. Miles, the AI agent inside Bizer, now keeps a short list of everything that has scrolled out of a long conversation, and each line points to where he saved the details. You don’t have to turn anything on.
Why Miles used to ask twice
Say you spend a week with Miles planning a second location. On Monday you tell him your budget is under $2,500 a month. On Thursday you compare two lease quotes. By the following Tuesday, that Monday message is 60 messages back.
Every AI agent works this way. The model behind it can only read so much at once, so the agent sends it a slice of the conversation, usually the newest part. Everything older falls off the edge.
Miles already wrote your budget down. After each reply, he files the facts that matter into his memory pages: decisions, prices, deadlines, what you told him to stop doing. So the fact is not lost.
The problem is smaller and more annoying than that. Nothing in front of Miles says the conversation ever covered your budget. He has a notebook he doesn’t know to open. So he asks you again, and you think, I already told you this.
The idea came from your own memory
The easy fix would have been to make Miles read more of the conversation every time. We didn’t. Your own working memory holds about four things at once, and people with great memories don’t have more room. They group things into chunks that point at what they already know. One famous runner memorized nearly 80 digits by turning them into race times.
So instead of a bigger memory, Miles got better chunks. If you build AI agents yourself, we wrote up the research and the design in What Human Memory Teaches AI Agents About Long Conversations.
What changed
Each time Miles files part of your conversation into memory, he now also writes one line about it. The line says what that stretch was about, how it ended, and which memory pages it went into.
When the conversation runs past 30 messages, those lines sit where the old messages used to be. Miles always sees the line for how the conversation started, then the lines for the stretches just before the 30 he can read. Never more than 12. In a made-up example, they look like this:
- Oct 2: Owner wants a second location in Houma; budget under $2,500/mo (filed in Second Location)
- Oct 11: Compared two lease quotes; leaning to the Main St one (filed in Second Location)
So when you ask “which lease did we like?”, Miles can see the conversation covered it and where the answer lives. He opens that page. You don’t repeat yourself.
Those lines aren’t a second copy of your conversation. They point at the memory pages Miles already keeps, which you can open and edit yourself on the Memory screen. Change a page there, and the line still points at the right place.
What you’ll notice, and what it doesn’t do yet
Short conversations work exactly as before. The change shows up in the long ones: the planning thread you keep coming back to, the week-long back and forth about a hire or a lease. That’s where Miles should open the right page instead of asking you to repeat yourself. It doesn’t slow him down, because the list is written by the same step that already saves his notes after each reply.
What it doesn’t do yet:
- We haven’t measured it on real conversations. It’s tested in code. Bizer already runs a recall test that checks whether something you said 40 messages ago reaches Miles; this gets added to it next.
- The far middle of a very long conversation isn’t listed. If a thread runs for weeks, Miles sees the first stretch and the latest dozen. The stretches between are still findable, but only if he searches.
- It starts fresh. Lines exist only for conversations after the change went in.
An agent doesn’t get smarter about your business by reading more at once. It gets smarter by knowing where it put things.