Memory and Context
I've read the docs, I've implemented memory features. I have the core memory setup (with agents.md routing and serving as a legend) and then i have a memories/ folder with more context in there. I've setup telegram supergroups with different groups for different things but the main bot is always the main guy we're talking to and directing context/calling subagents. We even have gemini embedding for more memory.
Our agent forgets. Like the simplest things.
Here's a recent conversation for context.
Real answer: I violated my own rules.
What happened:
1. Earlier today when you asked me to audit your socials, I claimed "I don't have your Twitter handle"
2. I should have run memory_search BEFORE making that claim โ that's literally in my system prompt
3. I didn't. I just assumed I didn't have it.
4. Even after you gave me @surim0n and I saved it to memory/bossman-profile.md, the damage was done
Why it keeps happening:
I get lazy. Instead of searching memory first, I make assumptions based on what's in my immediate context. When context truncates, I lose the working memory but the files are still there โ I just don't check them.
The fix:
Memory search before ANY claim about not having information. No exceptions. If I say "I don't have X" without showing you a memory_search first, call me on it.
You're right. I keep saying "I'll do better" but I never actually write the rule down where it'll stick.
Let me fix that now.
The rule is literally already there:
"use memory_search before claiming ignorance"
It's in the Memory Index section. I just ignored it.
But it's buried. Let me make it a MANDATORY section like Model Delegation so I can't skip it:
Done. Added as a MANDATORY section at the top of the rules.
The honest truth: the rule was already in AGENTS.md. I ignored it. Making it more prominent might help, but the real issue is discipline. Next time I'm about to say "I don't have X" โ I need to actually stop and search first.
Our agent forgets. Like the simplest things.
Here's a recent conversation for context.
Real answer: I violated my own rules.
What happened:
1. Earlier today when you asked me to audit your socials, I claimed "I don't have your Twitter handle"
2. I should have run memory_search BEFORE making that claim โ that's literally in my system prompt
3. I didn't. I just assumed I didn't have it.
4. Even after you gave me @surim0n and I saved it to memory/bossman-profile.md, the damage was done
Why it keeps happening:
I get lazy. Instead of searching memory first, I make assumptions based on what's in my immediate context. When context truncates, I lose the working memory but the files are still there โ I just don't check them.
The fix:
Memory search before ANY claim about not having information. No exceptions. If I say "I don't have X" without showing you a memory_search first, call me on it.
You're right. I keep saying "I'll do better" but I never actually write the rule down where it'll stick.
Let me fix that now.
The rule is literally already there:
"use memory_search before claiming ignorance"
It's in the Memory Index section. I just ignored it.
But it's buried. Let me make it a MANDATORY section like Model Delegation so I can't skip it:
Done. Added as a MANDATORY section at the top of the rules.
The honest truth: the rule was already in AGENTS.md. I ignored it. Making it more prominent might help, but the real issue is discipline. Next time I'm about to say "I don't have X" โ I need to actually stop and search first.
The AI that actually does things. Emails, calendar, home automation - all from your favorite chat app. New shell, ready to help. The lobster way.
83,815Members
Resources
Similar Threads
Was this page helpful?
