Knowledge

Produce it once. Consume it anywhere.

Provon is agnostic to what the knowledge is for. Learning Goals turn traced conversations into reviewed Knowledge Items, which any coding agent then pulls as Markdown or finds on demand. You decide what is worth learning; Provon handles production and consumption.

Knowledge overview Knowledge CLI
Learning Goal active

Output format preferences

Extract explicit preferences about response format, length, or tone.

Knowledge item

Lead with a short summary, then add details.

94% confidence · evidence conv_910f · tr_55a2
provon knowledge

$ provon knowledge pull

.provon/knowledge · 3 goals · 12 items

INDEX.md

goals/output-format-preferenc-1a2b3c4d.md

manifest.json tracks every generated file

$ provon knowledge find "summary first" --format paths

.provon/knowledge/goals/output-format-preferenc-1a2b3c4d.md

From traces to reviewed memory

Produce. Review. Consume.

  1. Define a goal

    01

    A Learning Goal states what to retain; there is no fixed category taxonomy.

    $ provon knowledge goals create "Output format preferences" "Extract explicit preferences about response format, length, or tone."

  2. Extract

    02

    Settled conversations run against every active goal; backfill past traces on demand.

    $ provon knowledge settings set-extraction true

  3. Review

    03

    Each item keeps content, a confidence signal, and evidence refs to its source.

    $ provon knowledge items get <item-id>

  4. Consume

    04

    Pull active items as managed Markdown, or match them deterministically.

    $ provon knowledge pull

Scope

Provon only does two things.

Produce

Turn traced conversations into reviewed items.

provon knowledge goals create Declare what is worth extracting
provon knowledge settings set-extraction true Extract from new conversations
provon knowledge goals backfill <goal-id> Extract over historical traces
provon knowledge items update Correct or archive against the evidence

Consume

Read reviewed knowledge into any agent or workflow.

provon knowledge pull Write .provon/knowledge Markdown + manifest
provon knowledge find "<text>" Deterministic match: json, md, or paths
provon knowledge status Verify the managed snapshot is intact
datasets: teach_project_knowledge Feed reviewed items into a dataset

Guarantees

Reviewable, predictable, and scenario-agnostic.

Goal defined, not categorized

Meaning comes from each goal instruction rather than a built-in taxonomy.

Evidence linked

Every item links back to the conversation and trace spans that produced it.

Deterministic to read

Find requires every term to match, returning predictable items or file paths.

Deduplicated

Repeated or similar evidence updates one item instead of stacking duplicates.

Produce knowledge your agents can consume.

Create one focused Learning Goal and pull the first snapshot.

Knowledge quickstart