Proposed Knowledge is where a person decides what the local model is allowed to teach your agents.

Pathrule's [local model](/docs/intelligence/local-model) can read a memory and work out what it says operationally: a problem and the fix that resolved it, a preferred choice, an ordered procedure. That reading is an inference, and an inference is not knowledge until someone vouches for it. Everything the model proposes lands in this queue first, and nothing in it reaches an agent until a person approves it.

## Where to find it

In Pathrule Studio, choose **Proposed Knowledge** in the activity bar. The panel has one tab per kind of advisory knowledge, each with its own queue:

| Tab | What it holds | What a proposal shows |
| --- | --- | --- |
| Remedies | A problem and the fix that resolved it | When, Do, and how strongly the source claimed the fix worked |
| Preferences | A preferred choice under a condition | When or context, Preferred, Instead of, and its strength |
| Procedures | An ordered workflow | The operation and its ordered steps |
| Context | A fact that changes the right approach | The fact, its topic and whether it is current |
| Rationale | Why something works the way it does | What it explains and the reason |
| Precedence | Which of two pieces of knowledge wins when they conflict | The winner, the loser and the condition or scope |

Each queue lists the oldest proposal first, so the queue drains rather than letting new items jump the line. The kinds are the advisory half of the [eight primitives](/docs/intelligence/knowledge-ir); CONSTRAINT and CHECK come from rules and never pass through this queue.

## Looking for proposals

On the **Remedies** tab, press **Look for proposals**. The panel shows "Reading your knowledge…" while the local model reads a batch of the workspace's most recently updated memories, newest first, and then a short summary of how many it read and how many it proposed.

A few properties hold every time:

- **It runs only when you press it.** Nothing schedules it, nothing watches your files for it, and nothing runs it when Studio starts.
- **It reads your memories, nothing else.** Not your source files, not your git history, not your rules or skills.
- **It runs on your Mac.** The model listens on this machine only, and the memories it reads do not leave it.
- **It does not repeat itself.** A memory that already has a proposal, pending, approved or rejected, is skipped. Pressing again costs only a readiness check, and a proposal you rejected is never offered again.
- **It reports outages as outages.** If the model is not installed or does not answer, the panel says "The local model did not answer" and offers to try again. It never turns a missing model into "nothing found". When individual memories could not be analysed, the summary says how many.

Today the button looks for remedies, the corrective kind. The other tabs use the same review, with the same decisions, for their own kind.

## Reviewing a proposal

Open a proposal to see everything the decision rests on:

- **The knowledge itself**, in the source's own words: for a remedy, the condition under "When" and the corrective action under "Do". A part the source did not quote is labelled as missing rather than filled in.
- **Where this came from.** The evidence state of the source (written by a person, worked more than once, worked once, untried, tried with an unknown outcome, did not work, or replaced), the source memory, and **Open the source**, which takes you to that memory so you can compare.
- **Authority.** Every proposal reads "Inferred, not vouched for". That is the honest status of anything a model produced, and it changes only when a person approves.
- **Warnings** where the extraction looks weak: a condition and an action that quote the same text, an action stitched together from short fragments, or evidence that is document structure such as a heading rather than knowledge.

The footer states the consequence before you click, because approving is not the same as delivering. It reads "Approving makes this available to agents" when the proposal can be delivered, or "Evidence is too weak to deliver, even once approved" when it cannot. A remedy whose source never claimed the fix worked stays undelivered even after approval, and a context fact that is no longer current is not delivered either.

Then decide:

- **Approve** makes it part of your knowledge, recorded with who approved it.
- **Reject** keeps the decision, so the same evidence does not come back as a new proposal.

## What approving changes

An approved atom becomes part of what Pathrule compiles for your agents, but it does not replace the memory it came from on its own.

At delivery, the [completeness gate](/docs/intelligence/safety) takes every memory with approved atoms and reads the memory as written against the knowledge those atoms' lines carry. When the lines represent the memory well enough, the memory is delivered in its compiled form on every channel its full text would use. When they do not, for example because the approved line kept what to use and lost what to avoid, the memory keeps arriving as written and its lines are withheld. Reading a memory by name always returns its full text either way.

[Agent IR](/docs/intelligence/agent-ir#how-a-memory-reaches-the-agent) describes the channels in detail.

## Where proposals live

Proposals and decisions are stored on your Mac, with the workspace's local store.

- **A local workspace** keeps them alongside the rest of its knowledge, and compiled delivery applies on every channel: the hook channels, the instruction files and Studio turns.
- **A cloud workspace** works too. The local model reads the copy of the workspace's memories Studio keeps on your Mac, and the proposals and decisions stay on that Mac. Compiled delivery applies to Studio turns on that Mac; agents started outside Studio on a cloud workspace receive memories as written.

Nothing in this queue is uploaded, and approvals do not sync to teammates. Each person decides for the turns their own Mac runs.

## What it never does

- It never approves anything on your behalf, and nothing it proposes reaches an agent before a person approves it.
- It never runs in the background or on a schedule.
- It never reads your repository. The only input is the memories your team wrote.
- It never edits the memory it read. The memory stays exactly as written, and remains what an agent receives whenever the compiled form is not good enough.

For how proposals are made, see [compilation](/docs/intelligence/compilation). For what happens after approval, see [Agent IR](/docs/intelligence/agent-ir) and [safety and fallback](/docs/intelligence/safety).