Seven ways to state the problem — written for people who already build and operate agents. No "agents forget everything" strawmen; every agent in production has memory infra. These name what's still broken anyway. Read each, score how hard it hits.
| learns from ↓ / teaches → | Human | Agent |
|---|---|---|
| Human | WORKS Slack, meetings, docs, shadowing, "how'd you close that deal?" Millennia of infrastructure. |
MANUAL OR LOST Deliberate teaching is prompt stuffing and bespoke RAG. Implicit teaching (every correction, edit, and rejection in live use) evaporates at session end. |
| Agent | DIES IN LOGS Agent discovers something mid-run. Insight lands in a trace nobody reads. |
NONEXISTENT Each agent's memory is private to its stack. Learnings never cross agent, framework, or team boundaries. |
Concrete version: your East Coast sales agent learns "enterprise buyers there want a security review before the demo." Your SF agent hits the same wall next week and learns it from scratch — at inference cost, maybe wrong. Your SF humans would have heard about it over lunch.
Every agent you deploy has memory — its own. LangGraph checkpoints here, OpenAI threads there, pgvector for that one custom app. N agents means N private memory stacks. Nothing crosses agent, framework, or team boundaries.
Who feels it: platform teams running agents across multiple frameworks and vendors.
RAG solved retrieval. Nobody solved the write path: what's worth persisting, when to update versus append, how to resolve a contradiction with what's already stored. So memory becomes an append-only log — grows forever, retrieval quality degrades, nobody trusts it.
Who feels it: infra engineers who've watched their vector store rot in production.
You log every transcript and embed every doc. That's storage. Learning is distilling "security review before demo works for East Coast enterprise" out of 500 transcripts — and surfacing it at decision time. Your stack does the first, not the second.
Who feels it: teams with big embedding pipelines and mediocre answer quality.
A human corrects an agent's output — the correction lives in that one thread's context. Next session, sibling agent, same mistake. And corrections are just the loudest case: every edit, approval, and rejection in live use is the same class of training signal, thrown away at session end. There's no promotion path from a one-off fix to knowledge every agent inherits.
Who feels it: anyone operating agents in production with humans in the loop.
Memory scoped to a user or a thread — solved. Memory scoped to an organization — with permissions, sharing rules, who-can-read-what — nobody built that. That's why your SF agent can't use what your East Coast agent learned, even though both learnings sit in a database somewhere.
Who feels it: platform and enterprise buyers thinking about agents at company scale.
Human knowledge lives in Notion, Slack, Drive. Agent memory lives in vector stores and thread state. Different formats, different permissions, ad-hoc ETL between them. A human learns something — agents don't get it. An agent learns something — humans never see it.
Who feels it: leaders deploying agents alongside human teams and expecting them to compound.
Every team: pgvector + a summarization cron + custom dedup + staleness hacks. Undifferentiated heavy lifting, rebuilt per app, interoperable with nothing. Agent memory infra is this cycle's "everyone rolls their own auth."
Who feels it: engineering leads who've now built this twice and dread the third.