AI gets better → humans get lazier → prompts get shorter → output gets generic.
When everyone's using the same models with the same low-effort prompts, everything starts to look the same.
The only differentiator left? An AI that actually knows your context. Your history. Your preferences.
Memory becomes the moat.
if you're building an AI product, here's the only benchmark that matters.
i call it the Human Refinement Benchmark: how much does a human need to fix the output before it's meets my expectations?
0 = absolute trash, doing it myself would've been faster
10 = one shot, perfection
my current scores:
@claudeai : 8
@ChatGPTapp : 6
@ManusAI : 6
@descript : 4
anything below 5, the product is probably costing you more time to fix.
what would you score the AI tools you use every day?
@paulg That's becoming the defining startup question.
For years it was "How do we use AI?" Now it's "What survives if AI gets dramatically better?"
Very different mindset.
@eglyman Every accounting firm I know is fighting the same battle: more work, same team.
This feels like a glimpse of what the next generation of firms will look like.
Our first Stanford hackathon...was wild, unscripted, hectic. It was our first hackathon ever, actually.
We didn't know what to expect, if anyone would even show up. People did. And they hacked. And the demos were 🔥
Our long-term memory API is out, use it, break it, build whatever you want with it. Tag @XTrace_ai with what you built and you may just win $1.
Docs below...
Every team is plugging in agents. Almost none of them are giving those agents real memory of the business.
So the agents hallucinate, repeat themselves, and start every conversation from zero. We don't onboard humans that way. Stop onboarding agents that way.
Half of management is just slow API calls between human brains. That existed because there was no other option.
There is now. If your company is queryable and artifact-rich, you can collapse entire layers that exist purely to route information.
thanks! we have a different schema for storing and retrieving memory, especially around artifacts. also you can use it with our private vector db which doesn’t exist for other solutions.
tldr; it’s encrypted, better for work-related use cases and long running agent tasks. we talk about it here in our research: https://t.co/Gfn4J48FT3
We just launched the new way of building memory for your AI agents.
Add long-term memory to your agent in one HTTP call.
No vector store to stand up, no dedupe logic, no session state to babysit.
Built on our open-source SDK (xmem):
1. extracts facts, episodes & artifacts from conversations
2. AGM-style belief revision — corrected info gets flagged "superseded," not piled up as noise
3. Postgres/pgvector + Redis under the hood
Free to start, go break it. Links in thread.