Every default AI chat has the same limitation: it knows nothing about your company until you tell it, and it forgets everything the moment the conversation ends. For a one-off question, that's fine. For actually running a business with AI as a working part of it, re-explaining your context every single time is the single biggest drag on making it useful.
I built a real system to fix this for my own setup, and the underlying idea is simple enough for any construction SME to copy without needing an enterprise platform.
The three layers that make memory work
Durable facts about how you operate. Things that rarely change: your company structure, your standard processes, the terminology specific to your business. This is the layer that stops you from re-explaining "this is what we mean by a technical office" in every new conversation.
A running log of what's actually happening. Recent projects, decisions made, open items. This layer changes constantly and needs to stay current — stale memory is worse than no memory, because it leads the AI to confidently reference something that's no longer true.
An identity/voice layer. How you want the AI to communicate, what tone fits your business, what it should and shouldn't do on its own without checking with you first. This is less about facts and more about behavior.
Keeping these three separate, rather than one giant undifferentiated memory blob, matters in practice: the durable layer barely needs updating, while the running log needs constant refreshing — mixing them makes both harder to maintain.
How I actually built this
My setup uses a real AI agent (Hermes, running continuously rather than only when I open a chat) with these three memory layers as actual files it reads and updates. When I tell it something worth remembering, it writes it to the right layer itself. When it starts a new conversation, it reads that memory back in before responding — so it already knows my company's context without me re-explaining it.
The mechanism doesn't need to be that specific setup. The pattern — separate durable facts from current state from behavior/voice, and give the AI a real place to read and write that memory — is what matters, not the particular tool.
What construction SMEs can do without an enterprise system
You don't need custom software to get most of the benefit:
- A file-based approach works well if you're already comfortable with structured notes (an Obsidian vault, or even a well-organized set of text files): one document for durable company facts, one for current projects/status, one for how you want AI to communicate.
- A shared knowledge base — even a simple structured document repository — gives any AI tool you connect it to real company context instead of a blank slate.
- Start small. You don't need all three layers perfectly built on day one. Even just a single "About my company" document that you paste in or connect at the start of a session beats starting from zero every time.
The tradeoff: autonomy vs. oversight
The more an AI system can write to its own memory without you checking each entry, the more useful it becomes — and the more it matters that you occasionally verify what it actually remembered is correct, not just what it claims to have recorded. I've had cases where a system's self-report of what it did didn't match what it actually wrote to file. The fix isn't avoiding autonomy; it's checking the real output periodically rather than trusting the summary blindly.
Company memory isn't a nice-to-have feature. It's the difference between AI as a tool you have to re-onboard every conversation, and AI as something that actually accumulates context the way a long-term employee would.