Ground · Chapter 03
Learning organisation
A business that corrects its own facts and rules from what happens.
A business that updates its own facts and rules from what happens. When an outcome contradicts a fact, the business flags it, checks it and updates it, so the next agent works from the corrected version. No model retraining needed. Glossary ↗
Collecting more records does not by itself mean that a business has learned. An outcome has to be able to change the facts, rules or bets that guide the next decision. Otherwise the business can record the same surprise repeatedly without doing anything differently.
This connects two kinds of correction. A finding about the present belongs in the fact layer as a proposed fact. A finding that supports or challenges a bet belongs in the hypothesis tree. Treating both as undifferentiated memory makes it difficult to decide what the next agent should do with them.
Closing the loop
Give useful conversations a path into that review. With the appropriate permission, capture a meeting in an existing notes tool, keep the source beside its summary and extract proposed decisions or facts into an inbox. A recording establishes what was said; someone still has to judge what the business should rely on. Granola can help capture the record, while version history can preserve how an accepted statement changes.
When an outcome contradicts an approved fact, flag the contradiction, check it and submit a correction for review. When it challenges a hypothesis, attach the evidence to the claim rather than changing the claim silently. The next decision should be able to use the corrected record.
Agents may also propose changes to their skills and workflows. Those proposals need the same discipline: a stated problem, a proposed change and a person responsible for accepting it. A system that can modify its own process still needs a way to establish whether the modification helped.
Learning from the process itself
Audit the loops as well as their outputs. Each should run on schedule, leave a record of its work, stay within its budget and contribute to its goal. A weekly review should lead to a redesign or retirement when a loop repeatedly fails that test.
The claim that business learning can live in facts and rules without changing model weights remains contested. Some of the research below argues for learned memory and changes within the model itself. That evidence stays attached to the hypothesis, so that a useful operational pattern does not become a claim that other forms of learning are unnecessary.
Try the loop with one permitted record, a duplicate and a proposed correction. Trace the resulting claim to its source, keep an ambiguous proposal unapproved and recover an earlier text version. The practical proof is an observed outcome that leads to an accepted change, followed by a decision that uses the corrected record.
Tools and setup
Existing implementations: Granola's getting-started guide covers calendar-linked and ad-hoc meeting notes. GitHub Desktop supplies a graphical version-history workflow.
- Install Granola through its official site, sign in, select the correct calendar if needed, and review sharing/retention settings. Use a permitted test meeting or note; obtain needed participant permission before capture. Keep the source link alongside the summary and check it against the available transcript.
- Start with manual copy/export into the approved inbox; check whether it includes notes, transcript or both. Do not promise automated transcript export without checking actual integration support.
- In GitHub Desktop, create or add an appropriate text repository; review selected files, commit and publish privately only when authorized. Exclude secrets and restricted source records before the first commit. Confirm one harmless change appears in history and can be recovered.
Setup paths checked against linked documentation on 15 November 2026. Verify current account support during setup; reader integrations have not been tested here.
Work through this with your agent
Paste the request into your assistant. It will interview you and help produce the working artifact.
Interview me to set up a capture-to-review loop for notes, calls and outcomes. Ask at most three questions per turn and wait. Ask what gets captured, whose permission is needed, retention/deletion rules, sensitive content, reviewers and useful destinations. Reuse current tools. Design a raw inbox with source IDs and timestamps, then extraction into proposed facts, decisions, stories and tasks. Keep originals and provenance; flag contradictions. Configure only approved destinations. Put suitable text and changes in version history, excluding credentials and restricted recordings; remember that deletion from a working file does not erase repository history. Test one permitted sample, a duplicate and a proposed correction. Save CAPTURE-RUNBOOK.md and a review record. Do not enable recording or background intake without specific approval. Completion means demonstrated capture, retrieval and review with an honest manual fallback if triggers are unavailable.
Use these documented setup paths as references; check current account support before acting:
https://docs.granola.ai/help-center/getting-started/granola-101
https://docs.github.com/en/desktop/overview/creating-your-first-repository-using-github-desktop