Lab memory in Slack — so no experiment starts blind, and nobody has to keep an ELN.
You talk to it as @Prior. Curie is the product; @Prior is the agent.
Labs already know what failed and why — it's in Slack. The day someone leaves, that knowledge is gone. Electronic lab notebooks are supposed to catch it; people don't file into a second system, so the dead experiment gets re-run.
In #experiments, mention a plan:
@Prior planning to fine-tune the ESM baseline, lr 1e-4, batch 32, v1
Curie checks Slack history, the Lab Record, and the literature, then returns one verdict:
- Collision — already tried (who, when, settings, outcome)
- Near-miss — close; here's what differs
- Clear — nothing found; go ahead
No forms. When a result lands in the channel (from a person or another agent), Curie updates the record. If a hypothesis actually flips, it says so without being asked. Ask @Prior where does the lab stand? for the map of what's supported, refuted, or still open.
Slack holds the memory (channel, List, canvases, App Home). Curie is a small Bolt app on EC2 — Socket Mode, no public URL. The model extracts parameters and writes the explanation; a rule decides the verdict so the same plan can't get two answers.
More detail: docs/architecture.md.
- Create a Slack app from
manifest.json, install it, turn on Agents & AI Apps. - Copy
.env.sample→.env(bot token, app token, OpenAI key, channel id). - Then:
pip install -r requirements.txt
python -m seed.run # once
python app.pyOptional checks: python scripts/smoke_tests.py api · python -m eval.run (must show zero false collisions).
Always-on hosting: deploy/aws-ec2.md.
| Path | What |
|---|---|
listeners/ |
Slack handlers + channel poller |
pipeline/ |
Verdict, logging, hypothesis ledger |
tools/ |
Search, Lists, literature, cards |
llm/ + prompts/ |
One LLM client, validated JSON |
seed/ |
Demo lab story |
eval/ |
Calibration — fails on false collisions |
docs/ |
Specs + architecture diagram |
Apache 2.0 · Slack Agent Builder Challenge, New Slack Agent track