Reference design · Intelligence
Project Memory for Site Works
- Macro
- Macro MCP server
- Routines
- Team memory
- LiveKit calls
- EU hosting
The problem
The record of a battery-room retrofit is spread across client emails, site calls, the crew channel, the drawing register and the risk assessments and method statements (RAMS). Every Friday someone writes the client's status report from memory. A hold point that nobody wrote down is left out of it.
What I would build
Each site gets a Macro project. The client is set up as a CRM company and the crew gets a channel. The engineer files client email threads into the project, and Macro records and transcribes site calls. A Friday routine reads the week's emails, calls and documents from the original items through Macro's MCP tools. It searches for a list of site terms that the firm maintains. It drafts the status report as a Macro document, with an @-mention on every line that links to the source. Where Macro's nightly memory disagrees with the latest source, the draft flags the line. The routine then opens a send-email review card. The engineer corrects and approves the report, and nothing reaches the client before that.
The logic flow
Click any node to see what it does, what it passes on and where a person must approve.
What it returns
I would watch one number: how many lines in each approved report the engineer had to correct, and how many of those the memory had wrong. This is a reference design on Macro's open-source workspace, not a system I run, so there is no measured saving to quote. The tool names are Macro's own, and the run shown is illustrative.
Why it fits
A utility's asset manager can now ask her own AI assistant to summarise a week of emails from three contractors. She gets a tidy page back before any of their Friday reports arrive. A contractor used to sell the weekly status report as proof of diligence. Now the client can assemble a version of it herself. A contractor could respond by having a chatbot write that report faster. That keeps the habit that made the report unreliable. Somebody still rebuilds a week of hold points, outage bookings and variations from memory. A model now does the rebuilding, with more confidence and the same gaps.
The client's assistant cannot see the site record itself. That record is the crew channel, the call where the electrician said the isolation was not yet signed off, the drawing register and the revision of the method statement that is actually on site. So I would put each job in one workspace. Each site has a project folder, the client is set up as a company and the crew has a channel. The client threads are filed beside them, and the site calls are recorded and transcribed. Every Friday a routine reads the week from the original items, not from the nightly memory. It searches for the exact words a site uses: hold point, isolation, outage, variation, defect. It drafts the report with a link to every item it cites. Where the memory and the latest source disagree, the draft says so. The engineer reads the draft, corrects it and approves the send. Calls about price stay out of the crew's shared memory.
This could serve a five-year strategy of selling the record behind the report, not the report. I offer that as a sketch, not a plan. Suppose every status line points back to the email, the call or the drawing it came from. The contractor can then give a client the evidence behind a claim, not a summary of it. After a few dozen jobs, that evidence answers the questions clients ask at practical completion and in a dispute. A tidy page written from memory answers none of them. The client's assistant can write a report, but it cannot supply the evidence.
See it run
Want one of these?
Every part of it is open source and you can host it yourself. I bring the build, the checks around the project's gaps and the judgement about where your gates should go.