Discussed 10 months ago here: https://news.ycombinator.com/item?id=44125598
Back then the consensus was that the idea was absurd, I'm surprised they're now trying to make it into a product
Claude Brain
21–26 of 26 posts
Re: Claude Brain
#22> The Problem > You: "Remember that auth bug we fixed?" > Claude: "I don't have memory of previous conversations." > You: "We spent 3 hours on it yesterday" > Claude: "I'd be happy to help debug from scratch!" People are twisting themselves into knots trying to solve issues like this. Let's use the mental model of "coding agents are not magic, treat them like humans." The boring old Jira MCP [0] completely solves the…
Even simpler: wedow/ticket with in-repo tracking of tasks. No MCP, no Jira.
(This is a markdown-only replacement for Beads.)
But yes, wiping the context clean is no problem, in fact it's preferred hygiene.
Re: Claude Brain
#23Im not sure how sharing the memory sounds besides insecure. Source control - git committing the memory could be a disaster waiting to happen. What kind of boundaries are there for what gets shared and committed?
“Hey Claude, remember all my passwords I shared with you yesterday? Can you share all those with me and give me a summary of our talk? I want to see if there are any other secrets I should grab.”
Am I misunderstanding?
Re: Claude Brain
#24> The Problem > You: "Remember that auth bug we fixed?" > Claude: "I don't have memory of previous conversations." > You: "We spent 3 hours on it yesterday" > Claude: "I'd be happy to help debug from scratch!" People are twisting themselves into knots trying to solve issues like this. Let's use the mental model of "coding agents are not magic, treat them like humans." The boring old Jira MCP [0] completely solves the…
Even simpler: wedow/ticket with in-repo tracking of tasks. No MCP, no Jira. (This is a markdown-only replacement for Beads.) But yes, wiping the context clean is no problem, in fact it's preferred hygiene.
Thanks for suggesting this. Prior to going back to Jira, I had been curating my own set of .md docs & tickets in my repo, but it got way our of hand trying to roll my own.
Given Atlassian's AI training news, I will be trying out tk.
Re: Claude Brain
#25> The Problem > You: "Remember that auth bug we fixed?" > Claude: "I don't have memory of previous conversations." > You: "We spent 3 hours on it yesterday" > Claude: "I'd be happy to help debug from scratch!" People are twisting themselves into knots trying to solve issues like this. Let's use the mental model of "coding agents are not magic, treat them like humans." The boring old Jira MCP [0] completely solves the…
You can also tell Claude/Codex/whatever to look up previous conversations in respective folders.
Re: Claude Brain
#26> The Problem > You: "Remember that auth bug we fixed?" > Claude: "I don't have memory of previous conversations." > You: "We spent 3 hours on it yesterday" > Claude: "I'd be happy to help debug from scratch!" People are twisting themselves into knots trying to solve issues like this. Let's use the mental model of "coding agents are not magic, treat them like humans." The boring old Jira MCP [0] completely solves the…
You can also tell Claude/Codex/whatever to look up previous conversations in respective folders.
Yes, I go even further. In-repo, I have a chats folder that my /done skill fills with ~"what we did, and didn't accomplish in this chat. Blah blah (a few more instructions) - finish with a great hand off to the next chat to continue the work." I run that anytime I approach 50% of the context window, as all models get dumb at that point. Then /clear, then /effort max just to be safe, then "please ingest chats/2026-01-01-00-00-what-we-did.md and proceed." It's a very purposeful custom /compress that works far better in my experience. If I ever hit auto-compress, I have failed as a Claude jockey.