@belbix

Founder of Tetu and Sacra

 Earth
Joined August 2017
  • Tweets356
  • Following87
  • Followers520
62Media
1/6 How we mess up our AI agents' memory "Why did you run tests for THREE HOURS?! Don't do that again!" You meant: stop wasting hours on endless checks. The agent might save: "Never run tests." Next task: "Why aren't the tests running?!"
1
1
35
more replies
5/6 Then you try to repair old emotional instructions with new ones. A blanket ban blocks a task, so you add an exception. Then another ban. I suspect this makes memory easier to clutter. The agent has to untangle it all just to do ordinary work.
1
5
6/6 Before deciding your agent is useless, open its memory and persistent instructions. Check which things you said in frustration became rules. Delete accidental bans. Rewrite useful rules to say exactly what is prohibited and when.
3
Predictions that advances in AI will make people work less are just as naive as expectations that more advanced weapons will make people fight fewer wars.
1
18
1/8 Your AI agent takes a task, goes silent, and leaves it unfinished I got burned by this with OpenClaw’s native harness and switched back to Codex. But even with Codex, agents still stop when I work through Telegram.
1
40
more replies
7/8 My default: avoid subagents and background execution unless they bring a real benefit. Parallelism can speed up a large job; otherwise it adds supervision. This is my OpenClaw/Telegram experience. I haven’t benchmarked other harnesses.
1
15
8/8 Put it in the task: “Work sequentially, without subagents or background execution. Wait for operations to finish, verify their results, and continue until the task is complete. If background execution is necessary, explain why and how completion will be checked.”
6
Are you sure your AI agent’s memory actually works? “Continue the task.” “Which one?” 😅 Or it remembers the task but forgets every decision behind it. My eight-harness comparison on LinkedIn: linkedin.com/feed/update/urn…
2
2
100
more replies
7/8 “Remember everything” costs latency, storage and money. Reindexing and retries add up. Paid embedding APIs are generally separate from Codex and Claude Code subscriptions. Worse: the agent may silently continue when retrieval fails.
1
8
8/8 You think the agent is using accumulated knowledge. It’s doing yesterday’s work again. Memory isn’t a five-minute setup. Don’t dive down that rabbit hole expecting one. How do you handle agent memory—and have you caught it failing?
10
Load more