hot take: agent memory is the wrong frame.
the real primitive is learned experience.
not:
“what did we talk about?”
but:
“what should I do differently next time?”
I just opensourced Aeonik Ingrain to explore this, starting with a simple @_HermesAgent sidecar.
it's a learned experience layer that turns past runs into source linked corrections, decisions, stale-plan warnings, lessons, and completed outcomes.
every run informs future behavior.
link below ⬇️
If your @_HermesAgent is broken on openai-codex with 'NoneType' object is not iterable, don't debug it. Paste this into Claude Code or Codex:
"ChatGPT's codex backend started sending output: null and the openai SDK doesn't null-guard for output in response.output: at _parsing/_responses.py:61. Patch that line in my Hermes venv (~/.hermes/hermes-agent/venv/...) to for output in (response.output or []):, then restart with launchctl stop https://t.co/zuw4nx2wUt.gateway && launchctl start https://t.co/zuw4nx2wUt.gateway. Re-apply after hermes update."
30 seconds. AI fixing AI. We live in the future and it's mostly broken.
Imagine every pixel on your screen, streamed live directly from a model. No HTML, no layout engine, no code. Just exactly what you want to see.
@eddiejiao_obj, @drewocarr and I built a prototype to see how this could actually work, and set out to make it real. We're calling it Flipbook. (1/5)
as i was falling asleep i gave claude code a drowsy goal and a 5-minute heartbeat and passed out
woke up to 20 commits. 4,800 lines of code. 182 passing tests. It specced, built, tested, found and fixed its own bugs. Overnight. Unattended. 🌙
a mini-harness in 3 files:
https://t.co/L905H4eTE0