The problem
Every agent you add starts with its own memory, so the phone line, the chat widget, and the document bot each learn your company from scratch and never share what they learned. The member who told the chat about the quiet floor three weeks ago tells the phone agent again tonight, and the back office never hears either.
What goes wrong
- Each agent keeps its own notes. Three agents means three versions of what the customer asked for, and none of them is the whole picture.
- Memory is a transcript dump. The agent reads everything it ever saw, gets slower, and still cannot say where a fact came from or when it stopped being true.
- The memory lives inside one vendor's model. When you want to change models, or export what you know, the memory stays with the vendor.
The approach
- Put one memory under every agent, per organization, that every surface reads from and writes to: phone, chat, documents, and the systems you run.
- Store each fact as a current understanding plus a dated trail of evidence, so every answer can say where it came from and when it last changed.
- The memory reads only what your team connects and shares with it. Admins can view, edit, and export every entry.
- Keep the memory separate from the model. The model is a setting you change. The memory is what your company owns.
When it works
When more than one agent touches the same customers, and the cost of asking twice is a guest or a client who feels forgotten. It needs a team that reads what the memory holds and corrects it, and an admin who decides what the memory can read.
Why context is the product
Most AI agents are fine at the sentence in front of them. What they lack is everything around it: who is asking, what they asked last time, what the house rules say, what changed yesterday. That surrounding material is context, and context is what makes an answer right for your business instead of right in general. The model generates. The context decides whether what it generates is useful.
A members club makes this obvious. A member calls at 9 p.m. about Friday. The phone agent needs the booking, the preferences from the last stay, the note the coordinator left in chat, and the rule about which upgrades need a manager. Four facts, four systems, one request. If the agent only has the transcript of this call, it answers like a stranger.
One memory, not one per agent
The fix is structural. Instead of each agent carrying its own notes, every agent reads from and writes to one memory for the organization. The phone line, the chat widget, the document intake, and the back-office workflow share it. When the chat learns that a member wants a quiet floor, the phone agent knows it on the next call, and the reservations workflow knows it when it places the hold.
Roo is the cockpit a company runs its AI crew from. The crew works inside the company's own systems, under its rules. The Brain belongs to the company. Early access February 2027.
What a good memory stores
Current understanding, and how we learned it
Each entry has two layers. The first is the current understanding: what we know right now, short and ready to act on. The second is the trail of evidence: dated entries, with the source document or conversation, showing how the understanding changed. An agent reads the first layer to answer and writes to both after it acts.
This is what lets an answer carry its own receipt. Quiet floor requested. Source: chat, 21 days ago. Your staff can check it in seconds. Your member never has to repeat it.
Entities, not transcripts
A transcript dump is not a memory. It gets longer, slower, and no more useful. A memory is organized around the things your company runs on: people, customers, projects, decisions, systems, procedures. A conversation feeds those pages. It is not the page.
Permissions and ownership
The memory reads only the sources your team connects and shares with it. If a folder is not shared, the agent cannot quote from it. The security page covers how we handle your data.
Your admins can view every entry, edit it, and export all of it. The memory belongs to your company, in a format you can read without us. If you leave, it leaves with you.
The model is a part you swap
Keep the memory separate from the model. The model is what reads the context and writes the reply, and the market for models moves every quarter. When the memory is yours and the model is a setting, changing providers is a configuration change, not a rebuild, and not a loss of everything your agents learned. The model will be replaced. The memory keeps what your agents learned. One memory with a source on every fact, separate from the model, fixes the three failures: repeated questions, facts nobody can trace, and memory locked to one vendor.
When it does not work
It buys little when one agent touches each customer. And it does not work if nobody reads what the memory holds and corrects it, because a wrong fact then repeats on every channel.
Where to start
Start with a straight answer. Ask Roo, Webaroo's AI: How would our AI agents share one memory of our customers? If there is a fit, the first paid step is the 10-day assessment. For a clinic or physician group, healthcare operations covers the patient calls and intake that one shared memory serves.
Book 15 minutes with Connor