If you're interested in how AI memory systems work more broadly, including provenance, trust state, and cross-tool portability that's what I'm building at
@statelessagent
https://t.co/0oVB4kuMxn
Anthropic shipped a full source map inside the Claude Code npm package. I read it. Here's how the memory system actually works at the source level. It's more thoughtful than most people assume. 🧵
Where it gets interesting: there's no provenance tracking. A memory doesn't know which conversation created it or what evidence supported it. Staleness is purely mtime-based. The system compensates with behavioral guardrails "the memory says X exists is not the same as X exists now" — but the gap is real.
There's also no automated cleanup. Old memories persist until you manually delete them. And the 200-file scan cap means active projects could eventually hit ceilings. These aren't dealbreakers, but they're worth knowing if you're a power user relying heavily on memory.
Practical tips from reading the source:
→ Keep MEMORY.md entries to one line, under 150 chars
→ Put detail in topic files, not the index
→ Write specific description: fields — it's your recall ranking function
→ Check ~/.claude/projects/*/memory/ to see what's accumulated
@AnthropicAI "each agent is dropped into a fresh container with no context and will spend significant time orienting itself"
this is the problem i've been obsessing over. just shipped something today that tries to help with it:
https://t.co/0oVB4kuMxn
Introducing Claude Opus 4.6. Our smartest model got an upgrade.
Opus 4.6 plans more carefully, sustains agentic tasks for longer, operates reliably in massive codebases, and catches its own mistakes.
It’s also our first Opus-class model with 1M token context in beta.
Introducing Claude Opus 4.6. Our smartest model got an upgrade.
Opus 4.6 plans more carefully, sustains agentic tasks for longer, operates reliably in massive codebases, and catches its own mistakes.
It’s also our first Opus-class model with 1M token context in beta.
Things are moving so fast.
introducing a concept i've been working on - Stateless Agent Memory Engine, or SAME
the problem: every ai session starts from scratch. you re-explain your project, paste in context, forget what decisions y'all made together. You spend 10 min setting back up and lose your focus.
SAME fixes that. it indexes your markdown notes locally and automatically surfaces the relevant stuff when you start a session. decisions get logged. sessions hand off to each other. your agent actually remembers.
how it works:
→ ollama embeds your notes on your machine
→ sqlite stores the vectors
→ your prompt gets matched to notes that matter
→ context shows up automatically
no files leave your laptop. fully local. Context goes to your AI tool the same as pasting it yourself.
works with claude code, cursor, windsurf — anything that speaks MCP.
built this in a week with opus 4.5. made it for myself but maybe it helps you too.
first thing i've shipped publicly. very nervous. feedback welcome. poke holes in it. <3
https://t.co/OwCqeUE0QX
Ready Freddie?
Retweet for your chance to win the Queen of all trips. The winner will get to see the London red carpet screening of @BoRhapMovie from #TMobileTuesdays. #contest
What a glorious haul we have for this week's #WinWednesday giveaway! Follow us and retweet for a chance to win a #SeaOfThieves journal, pencil, sticker set and pin badge! Get your entries in before 4PM GMT tomorrow. Good luck!