AI for Engineering Knowledge Management
PLM AI agents and product memory: why records are not memory, and how an AI intelligence layer turns engineering history into reusable knowledge.
·
⏱
8 min read
Dr. Maor Farid
Maor Farid is the Co-Founder and CEO of Leo AI, the first AI platform purpose-built for mechanical engineers. He holds a PhD in Mechanical Engineering and completed postdoctoral research at MIT as a Fulbright fellow. A Forbes 30 Under 30 honoree and former AI researcher, Maor leads Leo AI's mission to help engineering teams design better products faster.
BOTTOM LINE
PLM gives you a reliable record of what your company shipped. It was never designed to give you memory of why. Closing that gap is the difference between an archive you search out of obligation and a body of knowledge your team reaches for instinctively. An AI intelligence layer on top of your existing PDM and PLM turns scattered history into recallable product memory, so engineers reuse what works, avoid re-solving solved problems, and spend their time on the design decisions that are genuinely new.
Your PLM system knows what your company shipped. It rarely knows why. That gap is the reason engineers keep re-solving problems their colleagues solved years ago, and it is why the conversation around PLM AI agents and product memory matters now. A revision history tells you that a bracket changed from version A to version B. It does not tell you that the change happened because a supplier dropped a material, that two alternatives were rejected on cost, or that a field failure forced the redesign. Records preserve outcomes. Memory preserves reasoning. Most engineering organizations have invested heavily in the first and almost nothing in the second, and the cost of that imbalance shows up every time a new project starts from a blank screen instead of from everything the team already learned.
Records are not memory
A PLM platform is, at its core, a system of record. It governs revisions, enforces approval workflows, controls part numbers, and ties a bill of materials to released documents. That work is essential, and systems such as PTC Windchill, Siemens Teamcenter, and Arena PLM do it well. But a system of record answers "what is the current state" far better than it answers "how did we get here and what did we consider along the way."
Research on engineering organizational memory has made this distinction explicit for decades. A widely cited body of work on design rationale, published in CIRP Annals and Computer-Aided Design, notes that despite sustained academic and industrial effort to codify knowledge in IT tools, most of the knowledge in an organization still lives in the memory of individual engineers. The same literature warns that with greater personal mobility between companies, the people who hold required expertise are increasingly unavailable to consult, so teams must rely on knowledge captured independently of any one person's memory.
The practical result is familiar. The decisions that matter most, the trade-offs negotiated between design, manufacturing, sourcing, and quality, are made in meetings, emails, and conversations that never become structured PLM data. The record captures the answer. The reasoning evaporates. For a deeper look at why this happens, our piece on engineering knowledge management covers the organizational patterns in detail.
IN PRACTICE
The designs and answers we got from our external consultants were very preliminary. Today we get better answers, faster answers. A wider range of design directions that weren't available to us before. It feels like the world was opened.
Harel Oberman, CEO, Oberman Industrial Designs
What product memory actually means
Product memory is the ability to recall not just the artifacts an organization produced, but the context that produced them: the geometry, the specifications, the supplier choices, the rejected alternatives, and the rationale behind released designs. It treats your engineering history as a queryable body of knowledge rather than a filing cabinet of revisions.
This is different from data exchange. ISO 10303, the STEP standard documented in NIST primers, defines roughly 700 parts and does an excellent job moving product model data between systems, covering geometry, tolerances, materials, and classes of PDM information. STEP is about interoperability of data. Product memory is about retrievability of meaning. You can have perfect STEP files and still be unable to answer "have we ever designed a sealed enclosure rated for this environment, and what did we learn."
Real product memory has a few defining properties:
It spans the whole history, not just the latest release, so superseded designs and the reasons they were superseded remain reachable.
It crosses repositories, because knowledge is scattered across PDM, network drives, ERP, and specifications rather than living in one place.
It is searchable by intent and by geometry, not only by part number, so an engineer can find a prior solution without knowing it exists.
Our overview of PDM software for mechanical engineers explains where data management ends and where memory needs to begin.
The cost of forgetting
The absence of product memory is expensive, and the expense is measurable. The McKinsey Global Institute found that knowledge workers spend roughly 1.8 hours per day, about 9.3 hours per week, searching for and gathering information. For engineers the burden is more acute. A 2022 CADENAS survey of more than 100,000 engineers and designers found that nearly half spent at least an hour every day searching for parts.
That search tax has a direct design consequence: when engineers cannot find an acceptable existing part quickly, they create a new one. The U.S. Department of Defense has estimated the cost to introduce a single new part number at roughly $27,500 once tooling, qualification, documentation, and lifetime support are included. Multiply that by the duplicate parts a growing organization generates each year and the figure becomes serious. As design vaults grow and institutional memory fades through turnover, the share of redundant designs tends to rise rather than fall.
This is fundamentally a memory problem disguised as a search problem. Engineers are generally happy to reuse a proven component. They create new ones because finding the right existing part takes longer than starting fresh. The reasons engineers default to redesign are explored further in our article on why mechanical engineers dislike designing products.
How an AI intelligence layer delivers product memory
This is where Leo fits. Leo is an AI intelligence layer that sits on top of your existing PDM and PLM rather than replacing them. Your system of record stays the system of record. Leo adds the system of memory: natural-language and geometric search across a company's full engineering history, including CAD files, specifications, and past decisions. Integrations are available for SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter, Arena PLM, and other systems, so the knowledge already locked in those repositories becomes reachable.
The value driver is reuse before creation. When an engineer needs a part, Leo prioritizes parts the company already designed or already bought, then 120 million plus vendor options, before generating any new geometry. That ordering directly attacks the cost of forgetting. Industry analysis suggests engineers spend about 35 percent of their time designing parts that already exist, and finding the right existing part can cut reported BOM costs by roughly 15 percent.
Because Leo is trained on more than one million pages of engineering standards, books, and articles, it understands engineering intent rather than matching keywords alone. And it does this without compromising IP: Leo is SOC 2 certified and GDPR compliant, no AI is trained on your data, and your IP is never shared. To see how this connects to the broader idea of connected engineering memory, our post on BOM and product memory in engineering workflows lays out the vision.
Building product memory into your workflow
An intelligence layer only helps if engineers actually reach for it, so adoption matters as much as capability. The goal is to make recalling prior work the path of least resistance, easier than starting over. A practical rollout follows a clear sequence:
Connect the layer to your existing PDM, PLM, network directories, and ERP so it can see the full history rather than a single silo.
Make search the default first step of any new design task, querying by intent and by geometry before opening a blank CAD document.
Capture decision context as work happens, so the reasoning behind a choice is preserved alongside the released artifact.
Measure reuse rate and search time, and treat improvement in those numbers as a design-efficiency metric, not an IT metric.
None of this requires ripping out your governed processes. The PLM keeps controlling revisions and approvals; the intelligence layer makes everything that PLM has ever held actually usable. For a wider view of how these tools compound, see our roundup of engineering productivity tools. The organizations that treat memory as infrastructure, not as an afterthought, are the ones that stop paying the forgetting tax.
FAQ
McKinsey Global Institute, research on knowledge workers spending about 1.8 hours per day searching and gathering information, supporting the cost-of-forgetting estimates.
CADENAS 2022 survey of more than 100,000 engineers and designers, supporting the claim that nearly half spend an hour or more daily searching for parts.
NIST primers on ISO 10303 (STEP), supporting the description of STEP's scope, its roughly 700 parts, and its focus on product data exchange rather than knowledge.
Design rationale and engineering design memory research in CIRP Annals and Computer-Aided Design (ScienceDirect), supporting the distinction between records and reasoning and the loss of knowledge through turnover.
Turn PLM records into product memory
See how Leo makes your engineering history searchable and reusable.
Leo adds natural-language and geometric search on top of your PDM and PLM so engineers reuse proven work first. Book a demo.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
