Guide

Knowledge Extraction

Knowledge extraction is how the session stops being a one-off.

Source: docs/knowledge-extraction.md

Knowledge Extraction

Knowledge extraction is how the session stops being a one-off.

Review, dissent, and salvage produce raw learning. If that learning stays in chat, the next run starts cold. If it becomes hidden memory without review, it becomes unaccountable policy.

The useful middle is a durable, inspectable artifact.

What to record

Use the smallest artifact that changes future behavior.

ArtifactUse whenWrites to
MetisYou learned a situated pattern that should inform future work..oh/metis/<slug>.md
SignalYou found a measurement that indicates movement or risk..oh/signals/<slug>.md
GuardrailSomething must not happen again..oh/guardrails/<slug>.md
Outcome updateStatus, mechanism, or affected files changed..oh/outcomes/<slug>.md
ADRA decision now constrains future architecture.docs/ADRs/<NNN>-<slug>.md

This follows the shape of the RNA record skill: metis, signal, guardrail, outcome update, and ADR.

Meeting notes: summary is not the finish line

A transcript summary can still leave the organization confused. It may list topics and action items while missing the thing that matters: how the conversation changes aims, projects, guardrails, risks, decisions, and future work.

Model-fit framing gets the notes grounded. Knowledge extraction decides what survives after the meeting note has done its immediate job.

Start with the grounded note from docs/model-fit.md:

ContextWhy it matters
transcriptraw meeting evidence
attendees and rolesauthority, ownership, and follow-up context
active aims, projects, programs, and customer commitmentsthe big picture the note must tie back to
company vocabularynames and concepts the note should preserve
prior decisionswhat is settled versus reopened
known guardrails and risksconstraints the note should check against
desired output artifactthe shape reviewers need

Then extract only the pieces that should affect future work:

Meeting note contentDurable artifactWhy it survives
“We decided the onboarding assistant should not answer entitlement questions until the permission boundary is in place.”ADR or guardrailFuture implementation must respect the boundary.
“Customer-success notes show three teams asking for the same post-meeting project summary.”SignalThe demand pattern can be measured across accounts or meetings.
“The model confused program names when the transcript lacked project context.”MetisFuture note generation should require project vocabulary and active aims.
“The roadmap outcome now depends on the meeting-notes extractor.”Outcome updateThe outcome's mechanism or affected files changed.
“Someone should check whether Zoom summaries are good enough.”Usually no durable artifact yetIt is an open question until evidence or a decision exists.

The extraction pass keeps meeting notes tied to the big picture. A shallow note says what people talked about. A useful extracted artifact says what changed, what should constrain future work, and what evidence supports that change.

What not to record

Do not record:

  • generic advice;
  • things the model already knows;
  • unverified claims;
  • one-off preferences;
  • stale assumptions;
  • private or sensitive details that should not persist;
  • rules nobody has agreed to enforce.

Memory without governance is a liability.

Extraction questions

After review, dissent, or salvage, ask:

  • What did we expect?
  • What actually happened?
  • Why did the difference matter?
  • What future run should behave differently?
  • Is this a pattern, signal, constraint, outcome update, or architectural decision?
  • What evidence or provenance supports it?
  • Who can retire or override it?

Candidate to promoted artifact

Treat extraction as a promotion path: raw observation becomes candidate learning, candidate learning becomes a reviewed artifact, and reviewed artifacts become future context.

Not every observation deserves promotion.

A good metis artifact changes how a future agent acts. A good guardrail prevents a repeated failure. A good signal gives the next run reality contact. A good ADR names a decision that future code must respect.

Example: code review learning

ArtifactExample
Raw observationThe agent fixed duplicate notifications by adding a guard in the failing caller. Review found two other caller paths that could still send duplicates.
Metis candidateDuplicate-send bugs in this repo usually belong at the notification boundary, not individual caller paths. Caller guards suppress the visible symptom and leave parallel trigger paths exposed.
Guardrail candidateNotification duplicate prevention must be enforced at the send boundary or a documented equivalent. Caller-specific duplicate guards are not sufficient unless the caller is the only possible send path and review verifies that boundary.
Signal candidateA regression test that sends two events with the same idempotency key to the same recipient should produce one send and one duplicate-skip record.

How this connects to the loop

Loop stepExtraction question
IntentWhat aim should future work inherit?
Problem framingWhat constraint or landmine did we discover?
Solution searchWhich rejected path should future agents avoid repeating?
EvidenceWhat check should become reusable?
DelegationWhat role or tool boundary mattered?
VerificationWhat did review prove or fail to prove?
DissentWhat alternate failure mode remains live?
SalvageWhat learning survives after dropping the draft?

The learning is the asset. The artifact is how it survives.

Exercise

After a review or salvage pass, write one candidate artifact:

  1. Metis if the learning is a situated pattern.
  2. Signal if the learning is a measurement.
  3. Guardrail if the learning is a constraint.
  4. Outcome update if status, mechanism, or affected files changed.
  5. ADR if the learning constrains architecture.

Then ask whether a future agent should actually inherit it. When the answer is no, leave it as a note, not memory.

Go deeper