Back to Blog
Lean & Execution

Before You Ask AI Agents to Run Your Factory, Build a Usable Memory

Julian Zur-Lienen||4 min read
Before You Ask AI Agents to Run Your Factory, Build a Usable Memory

Everyone wants AI agents to close work orders, reschedule lines, push updates to ERP, and draft shift reports. Pilots start fast. Two weeks later the agent is guessing, people stop trusting it, and the trial goes quiet.

The problem is not the algorithm. The problem is the absence of a usable, shared memory.

What actually fails in agent pilots

Teams try to layer autonomy on top of fragmented truth. Work instructions sit as PDFs on a shared drive. BOM revisions live in email. Quality holds are in WhatsApp. Exceptions are in the head of the shift lead. The ERP has fields no one fills because no one is sure which field wins.

An agent cannot reason from what it cannot reliably find. If your best operator would need three people and five systems to get a correct answer, the agent will stall or invent one.

Documentation is infrastructure

For an agent to make a safe move, it needs to know:

  • What “done” looks like for the task
  • Where the source of truth sits, and who owns it
  • Which rule overrides another when there is a conflict
  • What the current version is, with effective date
  • The common exceptions, and what to do when they occur

If this is undefined or buried, you are not failing at AI. You are trying to automate a process that does not have a stable form.

Memory before autonomy

Start with a narrow slice. Pick one product family or one cell. Build a plant memory that a human and a machine can both use under time pressure.

  • List the top 20 recurring questions from the last month on that slice
  • Write each answer as a one‑page SOP with owner, revision, and next review date
  • Publish in a single canonical location with a stable link per SOP
  • Tag with process, variant, and equipment; record which document overrides which
  • Capture the 10 most frequent exceptions and the first safe action for each
  • Keep a decision log for unusual cases so the next operator does not start from zero

Do not over-engineer the tool. Clarity beats features. The rule is simple. One truth per topic. One owner. One place.

Add a retrieval layer, not an agent

Before you automate actions, make answers instantaneous. Stand up a retrieval system that lets people ask questions in plain language and gets results from your actual documents. Plug in SOPs, policies, change logs, NCR templates, and recent shift notes. Require citations so the user can click the source.

Host this knowledge inside the EU. Choose an EU company and an EU-only stack so your operational data stays under European jurisdiction and control. A data center on European soil run by a non‑EU company does not guarantee sovereignty. Under foreign law, that data can still be compelled. If your factory knowledge is going to power future automation, keep it legally protected at the root.

A 30‑day plan you can run now

Week 1. Inventory sources and questions. Pull the top 20 questions. Collect the relevant SOPs, forms, and policy emails into one folder. Name an owner for each.

Week 2. Clean and publish. Convert answers into one‑page SOPs. Set versioning, owners, and effective dates. Remove duplicates. Establish a single URL per document.

Week 3. Wire up retrieval. Stand up a private, EU-hosted retrieval system. Ingest the cleaned set. Turn on citations. Route unanswered questions into a backlog.

Week 4. Pilot on one line or team. Measure time to answer for common questions. Track rework caused by outdated instructions. Close the top five knowledge gaps from the backlog.

Two numbers will tell you if the memory works. Time to answer under 10 seconds for common questions. Close to zero rework caused by obsolete or conflicting documents.

When to introduce narrow agents

Add small, well-bounded automations only after the memory holds.

  • Auto-create an NCR from a template when a specific SPC breach is logged
  • Open a maintenance ticket when a sensor alarm matches a documented rule
  • Update an ERP attribute after QA sign‑off when the SOP says it is safe

All of these rely on clean definitions of done, ownership, and exceptions. If those are shaky, keep investing in the memory.

The strategic payoff

As software gets cheaper, execution speed decides outcomes. A shared, sovereign memory cuts decision latency, training time, and rework. It also makes future agents simple to verify, because every action traces back to a cited rule you control.

Build memory first. Autonomy will follow, and it will stick.

Want a pragmatic 30‑day Memory Sprint on an EU‑only stack for your plant? Talk to EUnexia.

Sources

  • 55-pathtoaiagent.md