LinkedIn’s Cognitive Memory Agent is basically an AI memory layer that keeps track of past context, facts, and workflows so agents stop acting like they just woke up.
“Persistent memory” sounds nice until you ask how they prevent memory poisoning and cross-user leakage. What’s their story on provenance and access control here—are they treating memory like untrusted input with strict scoping per tenant/user, or is it more “we’ll filter it and hope”? Because someone’s going to try to stuff that memory with garbage on purpose.
When they say “persistent memory, ” are they storing it more like an append-only ledger with provenance (who/when/where it came from) or is it basically a mutable blob they keep summarizing and rewriting over time? I might be wrong here.
