Memory is the new vendor lock-in.
Everybody is talking about agentic memory and they are right: agents that remember, that don't need you to explain things over and over are the next unlock.
If switching tools wipes your agent's memory, you're just renting it. 🧵
I told Junior to name my subagents HR-safe funny names and here's what I got:
- Lint Eastwood (working on a lint issue)
- People Pleaser Pete (working on optimizing people queries)
- Avatar Ava
- Stacktrace Stella
- Cache Me If You Can
- Professor Cache-A-Lot
I've added cache warming to Junior this week and today it was very interesting to observe 2 Anthropic outages where they wiped the caches. This triggered our breaker and sent an issue to our @sentry instance.
Total cost of the outage for us is a few hundred dollars.
I set Junior to talk to me in Star Wars lingo from The Clone Wars era and got this today:
>The Jedi Code says don't merge with known bugs, Master, but the PR is approved
Worth every token
storytime: when we were trying to sell mistle (slopware factory), we started off by doing pilots with several companies and we realized that most of their codebases were pretty messy and were from a time before LLMs.
agents generally adapt to existing patterns so it meant that if you were mocking everything from your mom to database calls, they'd do the same. if your codebase didn't have typechecking or linters, they'd happily continue writing code that's full of bugs that could've been caught out early.
our product was automatically nerfed into the ground before it even had a chance to prove itself. why? because it'd 100x the slopslinging and give them more work.
(also because of this, an inhouse build was almost always the better option because it could be tailored to the existing state of their codebase, team, processes, tools, data, etc.)
so yeah, if you've never made your codebase productive for humans to work on pre-AI, you're not going to see much difference with agents today.