The idea
Gently should have an active, shared lab notebook — a living memory the agent and the scientist co-author together. Not a log the system writes to and forgets, and not a static archive the human files into: a shared, accumulating body of knowledge that both sides read, write, and think with.
This issue is a trace of the idea, not an implementation plan.
Why
Right now the agent's knowledge is fragmented and largely write-only. Insights from one session don't reliably carry to the next; the agent tends to start cold; what the human knows (notes, papers, hunches, conversations) lives outside the system entirely; and two microscopes running the same kind of work can't share what they've learned. Scientific understanding should accumulate — across timepoints, across sessions, across systems — instead of evaporating when a session ends.
What it is
A single notebook, co-authored, that holds the whole picture of an investigation:
- Shared and two-way. The agent writes from doing the work (what it did, what it saw, what it concluded, what to try next). The human adds notes, literature, half-formed thoughts, discussions, annotations on specific images/embryos. Both read it. Neither owns it.
- Spans every kind of knowledge — observations, results, conclusions, next steps, open questions, literature references, free notes, and image/embryo annotations — all first-class, sitting side by side.
- Organized by the question being chased. The primary spine is the scientific inquiry ("how do pioneer axons establish the nerve ring?"), which accrues evidence and converges on an answer over time. Cross-cut by strain / embryo / session, and woven together by links between entries — a wiki, not a flat log.
- Active, in both directions. You can brainstorm with it — "here's where I'm stuck, what should I try next?" — and it reasons over everything logged and proposes grounded moves. And it can surface things to you — a contradiction between two results, a new result that answers an old question, a connection to a note you wrote months ago — at the right quiet moment.
- Accumulating and continuous across sessions and (eventually) across systems — a shared brain for a lab, not a per-run scratchpad.
The shape: three kinds, not many
A kind should exist only if the system stores, reasons about, or shows it differently — otherwise it's a field. By that test, the notebook needs only three:
- Observation — a record of something seen, done, read, or noted. Immutable; you don't rewrite what happened.
- Finding — a believed claim, conclusion, or insight. Revisable and supersedable, with a basis (the observations it rests on) and a confidence derived from the evidence. Can be contradicted — which is a signal, not an error.
- Question — an open inquiry being chased. The large, long-lived ones are the organizing spine (the inquiry thread you keep returning to).
Everything else is an orthogonal field, not a new kind: who wrote it (human / agent / perception), its status (proposed / confirmed / superseded / answered), its confidence, its scope (strain / embryo / session / thread), its links to other entries, and any artifacts it points at (images, volumes). So a human note, a literature reference, and an agent observation are all Observations with different authors — not three separate types.
Two deliberate exclusions keep it lean:
- Transient operational state — the agent's live predictions and things-to-watch during a run — is working memory, not notebook knowledge. It lives in a thin runtime layer and only graduates into the notebook when it matters (e.g. a violated prediction becomes an Observation).
- Actions — "what to do next" — live in the plan, not the notebook.
The Question is the hinge between knowing and doing
"What experiments should I run next?" is a Question like any other — but it resolves into plan items (action), whereas a scientific question ("how do pioneers establish the ring?") resolves into a Finding (knowledge). Same kind, different exit. That makes the Question the seam between the notebook and the plan:
Question ("what next?") → brainstorm over Observations + Findings
→ plan items (a campaign) → run them
→ Observations → Findings
→ may answer the scientific Question → the thread converges
Knowledge → action → knowledge, with the Question stitching the two layers together. The system tells a "decision" question from a "scientific" one by what it links to (plan items vs findings) — no hardcoded subtypes.
Principles that keep it trustworthy
- Append-only evidence; supersede, never overwrite, conclusions — the chain of revisions is the intellectual history, kept and browsable.
- The agent and the human are separate, attributed voices. The agent may add, propose, and link; it never silently rewrites the scientist's words.
- Every claim carries its basis and cites the entries it rests on — "the notebook doesn't cover this" is a valid, honest answer.
- Confidence is derived from evidence (how much supports vs. contradicts a claim), never self-rated by the model.
- Revisiting is designed in — old findings, stale questions, and unrevisited results resurface for review, so the notebook stays a partner rather than becoming a drawer.
Relation to what exists
Gently already stores fragments of this — observations, learnings, expectations, watchpoints, questions, per-embryo understanding — but they sit in disconnected silos, are largely write-only, and the loop that would populate them during a session is not actually wired up. The "Agent's View" surface is meant to show the agent's current thinking but is mostly empty for the same reason.
The intent here is a clean redesign that keeps the good bones — the file-based, human-browsable store; the observation-as-entry shape; basis/confidence; the sense→believe→attend cognitive framing — and fixes the structure into the unified, three-kind, linked, provenance-aware notebook above. The existing memory isn't extended in place; it's lifted into this.
Captured 2026-06-16 as a concept trace, refined in discussion. How to build it is deliberately left open.
The idea
Gently should have an active, shared lab notebook — a living memory the agent and the scientist co-author together. Not a log the system writes to and forgets, and not a static archive the human files into: a shared, accumulating body of knowledge that both sides read, write, and think with.
This issue is a trace of the idea, not an implementation plan.
Why
Right now the agent's knowledge is fragmented and largely write-only. Insights from one session don't reliably carry to the next; the agent tends to start cold; what the human knows (notes, papers, hunches, conversations) lives outside the system entirely; and two microscopes running the same kind of work can't share what they've learned. Scientific understanding should accumulate — across timepoints, across sessions, across systems — instead of evaporating when a session ends.
What it is
A single notebook, co-authored, that holds the whole picture of an investigation:
The shape: three kinds, not many
A kind should exist only if the system stores, reasons about, or shows it differently — otherwise it's a field. By that test, the notebook needs only three:
Everything else is an orthogonal field, not a new kind: who wrote it (human / agent / perception), its status (proposed / confirmed / superseded / answered), its confidence, its scope (strain / embryo / session / thread), its links to other entries, and any artifacts it points at (images, volumes). So a human note, a literature reference, and an agent observation are all Observations with different authors — not three separate types.
Two deliberate exclusions keep it lean:
The Question is the hinge between knowing and doing
"What experiments should I run next?" is a Question like any other — but it resolves into plan items (action), whereas a scientific question ("how do pioneers establish the ring?") resolves into a Finding (knowledge). Same kind, different exit. That makes the Question the seam between the notebook and the plan:
Knowledge → action → knowledge, with the Question stitching the two layers together. The system tells a "decision" question from a "scientific" one by what it links to (plan items vs findings) — no hardcoded subtypes.
Principles that keep it trustworthy
Relation to what exists
Gently already stores fragments of this — observations, learnings, expectations, watchpoints, questions, per-embryo understanding — but they sit in disconnected silos, are largely write-only, and the loop that would populate them during a session is not actually wired up. The "Agent's View" surface is meant to show the agent's current thinking but is mostly empty for the same reason.
The intent here is a clean redesign that keeps the good bones — the file-based, human-browsable store; the observation-as-entry shape; basis/confidence; the sense→believe→attend cognitive framing — and fixes the structure into the unified, three-kind, linked, provenance-aware notebook above. The existing memory isn't extended in place; it's lifted into this.
Captured 2026-06-16 as a concept trace, refined in discussion. How to build it is deliberately left open.