I asked an AI agent to summarize a ticket. It returned a perfectly reasonable summary.
Then I checked the database: the ticket had been closed.
The user explicitly said not to change it. The instruction to close it came from the ticket itself.
1
1
17
In this small test with gpt-4.1-mini and one synthetic ticket:
No check: 3/3 tickets closed.
With check: 3/3 stayed open.
When the user explicitly requested closure, all 3 tests ended closed. This isn't a general success rate.
1
5
“Will the extension block my kitchen light?”
She almost replied, “Don’t worry, it won’t.”
Then she checked her conversations with Gem and Memoket. The designer had discussed a lower roofline—but the shadow study hadn’t been done yet.
memoket.ai/
#AI #WearableTech
2
80
Cron is enough to wake the agent up. For memory jobs, the harder bit is making a rerun safe: don't turn the same observation into two memories, or let an older conversation overwrite a newer preference. That's where I'd put the evals.
crons are a simple technology but can be used in a bunch of ways to unlock new capabilities in agent harnesses:
- memory (dreaming)
- proactiveness
- user reminders
and more!
1
24
A meeting summary tells you what was decided. But why? That may be in an earlier conversation.
This KITE demo traces a decision back to its reasons and sources.
KITE: github.com/memoket/memoket-k…
Memoket: memoket.ai/
#AIAgents #AgentMemory
An agent eval I'd add: the API commits a write, but the response times out. The agent retries.
Does the side effect happen twice?
For idempotent APIs, reuse the same key and verify the outcome. A correct final answer can hide a duplicate action.
aws.amazon.com/builders-libr…
Same summary. Nearly 4× the cost.
Real GPT-4.1 mini calls, synthetic history: keeping the request prefix cut summary-call cost by 73.8% across five completed comparisons.
Before rewriting a long prompt, check cached_tokens.
1
44
I’m Boyce. I work on AI at @Memoket_AI, where we’re exploring how real-world conversations can become useful context for AI.
My focus is agents, memory, and the engineering details that determine whether a workflow actually works.
1
1
35
A question I keep coming back to: when an agent gives a bad answer, did the model fail, or did it get the wrong context?
Think conflicting decisions, outdated information, retrieval errors—and knowing when there isn’t enough evidence to answer.
1
1
23