Learn more about Types of AI Agents here →
AI agents need more than training data to solve real-world problems. Martin Keen breaks down four core techniques used in modern AI agents: Skills, MCP, RAG, and Memory. Learn when each approach should be used, how they work together, and why they play different roles in helping AI agents access information, follow procedures, and learn from experience.
AI news moves fast. Sign up for a monthly newsletter for AI updates from IBM →
AI was used in the creation of the transcript and metadata for this video.
#skills #mcp #retrievalaugmentedgeneration #memory #aiagents
source
@IBMTECHNOLOGY, which of these would an ontology supported by knowledge graph or RDF store sit? Or is something else entirely?
Agent Skills (2:18–3:54)
MCP – Model Context Protocol (4:14–5:48)
RAG – Retrieval-Augmented Generation (5:59–6:55)
Memory (7:03–8:13)
i actually created something that combines all the first 3, i wrote an MCP server that wraps around copilot 365 retrieval RAG api for sharepoint documents (like SysRS and PRD documents) and it has a skill that tells it how to call the api, where to look for etc.. it has worked so far EXTREMELY well, so much that in some cases i'd ask it why something isn't working right, it''d run to read the docs and implement the fix CORRECTLY.
it doesn't have to be vs. it can be all of the above.
i'm curious to know if githubcopilot has a concpet of memory.
Sorry, but I think this is unprecise and a bit fuzzy. A real-world MCP example would be better. And what exactly do you mean with "Memory"? Which technology are you talking about? Finetuning the model, Lora?
The content in this video was explained with a master-class skill, making it easy to internalize in our Memory. I feel like I am building a solid RAG-style understanding of AI, without needing to have MCP for any other learning site.
Martin always does a good explaining things. Of course, he's on Team IBM and a lot of people are involved but they have a good presenter on their team.
What video a do you have 4 digital literacy ?
Simple and effective, thanks for this content!
everyone keeps calling memory “learning,” but most agent memory is just stored context with bad retrieval and no real adaptation
.
Very good video. Thanks. Now what to consider for MCP from a security point of view? Thanks
Great video, thank you. But one thing I'm wondering is how do you prevent agent memory getting too large, polluting the context window?
Normally, if you want a robot to learn a new task, you may need to collect training data, retrain the model, and test it again.
ICL tries to change that.
Imagine you have a robot and say:
“Pick up this object, put it in the box, and then close the lid.”
Instead of retraining the robot, you give it instructions, examples, or demonstrations. The robot uses that context to figure out what to do right now.
So:
Traditional robot:
New task → collect data → train → deploy
ICL robot:
New task → show/instruct → understand context → act
The exciting part about Gen-1.5 is the possibility that robots can handle more complicated, unfamiliar tasks from context, rather than being trained specifically for every individual task.
But there's an important distinction:
ICL isn't necessarily permanent learning.
The robot may successfully adapt during the current task without actually changing its underlying model permanently.
If this technology works reliably, robots could become much more general-purpose.
Instead of manufacturing:
one robot for one task
you could have:
one capable robot + instructions/examples → many different tasks.
That's a major step toward general-purpose physical AI.
LLMs showed that AI can adapt to new problems from context. ICL for robots is trying to bring that same flexibility into the physical world.
Fantastic breakdown. Huge thumps up to these videos. Super valuable <3
If RAG provides the script ( the exact words and facts), what process trains the actor ( AI) to deliver those lines with perfect, professional execution?
When an AI scans five different policy documents 📄 to find a hidden contradiction , does that count as a basic data retrieval job, or are you witnessing the model use its own logic and context to figure it out?
Human knowledge constantly expires. When an AI searches a database and finds an old truth and a new truth , how does it know which one represents the current reality?
How about the ' Goldilocks ' problem of chunking? How engineers find the perfect size for a text slice 🍰?
A pie chart 📈 or a complex table is meant to be scanned with human eyes , not parsed by an AI database. How difficult is it for modern RAG systems to convert those shapes and columns into something a language model can read?BTW, I am also aware that Multimodal RAG is actually one of the biggest battlegrounds in AI development right now.
If an AI is handling a complex task , at what exact moment does it close the reference textbook ( RAG) and pick up a calculator or browser ( Tool Use)?
Thanks for the video. Very clearly explained!
A useful way to operationalize this distinction is to define ownership boundaries: Skills encode repeatable procedures, MCP standardizes tool access, RAG supplies relevant evidence, and memory preserves task-specific state. The key governance question is what may be written to memory—and when retrieved evidence should override it.
Thanks mate!
So this is not mcp vs skills vs memory vs rag but all together
Really like this video, though I’m not sure how to set this up in Cursor/Codex, also MCP is not clear to me
Skills are more like suggestions
how did the guy get so good at mirror writing?
This is actually a really good way to break it down. The whole AI agent space is moving so fast.
This is the kind of AI content I wish I’d found sooner. Clear, practical, and easy to follow without oversimplifying the topic. The explanation of how these pieces work together in real AI agents was especially helpful. Great job!
great video
It looks like MCP is only used for the "tools" part of it. It can however, provide "resources", too. Which could be interface to, or substitute for, RAG. I don't like RAG as it never picks the right document when you have lots of similar documents with little variations in it (e.g. manual pages of a software with different versions).
That 's Great. 💯
Since RAG relies entirely on human written knowledge, what happens when two different people write down conflicting facts in the company database, and the RAG system pulls both? How does AI decide who to believe?
WRT RAG, what happens to information that is never written down? If a team relies on un-documented know-how, how do we capture that for AI (without asking humans to write ✍️ manuals)?
Which technique is most effective at reducing hallucinations in factual reporting?