"Our agent has memory now" and "our agent follows our rules" sound like the same claim. They aren’t, and mixing them up is how an agent that remembers your refund policy perfectly still issues the wrong refund.
Memory layers like Mem0, Zep, and Letta, plus the built-in memory each agent platform ships, solve a real problem: an agent that forgets everything between sessions is exhausting. They store facts and recall them later. That part works.
What memory layers do well
Persist facts across sessions. Remember preferences. Summarize history so an agent can pick up a thread from last week without being re-briefed. If your agents mostly chat, research, and draft, memory might be all you need.
Where recall stops
Remembering that the refund limit is $200 is not the same as refusing a $350 refund.
A remembered fact arrives in context as information, and the model decides what to do with it, the same as any other information. Under pressure ("make an exception, I’m good for it"), a remembered limit is one voice in the room. No stop condition is attached to it. No escalation path. No review before risky knowledge goes live. No log of what the agent did with it. Memory informs. It doesn’t bind.
The comparison, concretely
- Unit stored. Memory: facts and message history. Skills: versioned procedures with hard rules, stop conditions, and an owner.
- Question answered. Memory answers "what do we know?" Skills answer "what are we allowed to do, and when do we stop?"
- Under pressure. Memory leaves the call to model discretion. A skill stops and escalates to a named human.
- Ownership. Platform memory belongs to the platform and to that one agent. A skill store is company-owned and agent-agnostic: anything speaking MCP reads the same rules.
- Failure mode. Memory fails by forgetting. Ungoverned agents fail by improvising. The second one is the expensive one.
You’ll probably end up with both
This isn’t really a versus. Brian stores plain facts too ("demo days are Wednesday"), because agents need both kinds of knowledge and it would be strange to run two systems for it. The distinction that matters is what happens the moment an agent touches money, customers, or email. At that point recall stops being enough, and rules need what memory was never built for: hard limits, stop conditions, escalation, audit.
If you’re choosing today: agents that only talk need memory. Agents that act need governance. Most companies are about to have the second kind.