# Correcting Connect from a conversation, end to end

Tell the Assistant it is wrong and the correction becomes a proposal to write a memory, showing the wording and the tier before anything is stored. Confirming writes it through the memory service. The proof it worked is not the acknowledgement in the chat — it is the fact appearing in Memory and the next answer drawing on the new wording instead of the old.

- **Status:** Available
- **Audience:** both
- **In the app:** #/home, #/memory
- **Last verified:** 2026-09-10
- **Canonical:** https://connectbyjbrh.com/docs/workflows/assistant-correction/

## The path a correction takes

1. Trigger — an answer is wrong, or out of date, and you say so in the same conversation.
2. User event — the correction is posted as an ordinary message; there is no separate correction mode.
3. Authentication and workspace resolution — the correction will belong to this workspace and no other.
4. Ingest — the correction joins the tab's conversation, where the wrong answer is still visible above it.
5. Canonical record — the subject is resolved: a fact about the business, about one person, or about one company decides which tier it belongs to.
6. Reasoning — the model drafts the durable wording, which is usually not your sentence verbatim: 'no, it is two days now' becomes a statement that will still make sense in six months.
7. Knowledge, memory and rules — the service decides the tier from the subject, and an existing statement that contradicts the new one is superseded rather than duplicated.
8. Autonomy and approval — the proposal is shown to you with its wording and tier. Nothing is stored until you confirm.
9. Action through the owning service — `remember` writes through the memory service; `forget` removes one that should no longer apply.
10. Result — the stored fact exists, dated, attributed and visible in Memory.
11. Relationship and memory — a fact about a person is attached to that person, so it reaches every channel rather than only the Assistant.
12. Audit and Needs You — who corrected what and when is kept; a correction that contradicts a Knowledge source is surfaced rather than silently winning.

> **Note** The wording matters more than people expect. A stored fact is read back into future prompts inside a budget of roughly 700 characters, so a correction written as a durable statement earns its place; one written as a reply to a specific sentence reads as noise later.

## Fact, preference or rule

Three quite different things get called corrections, and only one of them belongs in Memory. Getting this right is the difference between a correction that sticks and one that has to be repeated every few weeks.

| What you are correcting | Where it goes | Why |
|---|---|---|
| A fact about the business or a person | Memory, through `remember` | It is knowledge, and it should reach every channel |
| A document or a source that is out of date | Knowledge | Grounded answers prefer Knowledge; leaving it stale keeps producing the old sentence |
| How Connect should behave — tone, escalation, what never to promise | A standing instruction | Behaviour applies across channels and is not a fact to recall |
| A single wrong record | The record itself, in the grid or on its screen | The Assistant reads records; fixing the memory does not fix the row |

The fourth row is the one that catches people. If the Assistant says a customer's phone number wrongly because the number on the record is wrong, storing the right number as a fact leaves the workspace with two answers and no fix. Correct the record.

## Proving the correction took

1. Confirm the proposal and read the wording it stored.
   - Result: The memory appears in [Memory](/docs/memory/) with a date and the person who made it.
2. Ask the original question again, in a new tab.
   - Result: A new tab has no conversation to fall back on, so the answer has to come from the stored memory rather than from the correction sitting three messages above it.
3. If the old answer returns, look for a Knowledge source saying the old thing.
   - Result: That is a [source conflict](/docs/assistant/source-conflict/) and is reported rather than resolved by preference. Fix the source too.

Re-asking in the same tab proves almost nothing. The correction is in the conversation the model can already see, so a right answer there does not demonstrate that anything durable was written. This is the most common way a correction is believed to have worked when it has not.

## What a correction does not do

- It does not train anything. No model learns from your workspace, and no correction changes a model's weights.
- It does not rewrite history. Answers already given stay in their conversations as they were said.
- It does not reach another workspace. A stored fact is scoped like every other row, under row-level security.
- It does not override a record. Records are the truth about themselves; what is stored is what Connect knows around them.

What it does do is change the next answer, on every channel, for everyone in the workspace — which is why the tier and the wording are shown before it is written rather than after.

## Questions

### Can I correct something without the Assistant proposing a memory?

Yes — edit it directly in Memory. The conversational route exists because the moment you notice a wrong answer is the moment the correct wording is clearest, not because it is the only door.

### What if I correct the same thing twice with different words?

The service supersedes rather than accumulates, so the newer statement replaces the older one about the same subject instead of leaving two contradictory memories for a later answer to choose between.

### Does a correction apply to phone calls too?

Yes. Memory is assembled into the brief a call is given, within its own character budget, so a fact corrected in the Assistant is available to a voice conversation that starts afterwards.

## Related

- [Correcting a fact without leaving the Assistant](https://connectbyjbrh.com/docs/assistant/assistant-memory/)
- [Memory in Connect](https://connectbyjbrh.com/docs/memory/)
- [Knowledge in Connect](https://connectbyjbrh.com/docs/knowledge/)
- [When the Assistant finds two answers](https://connectbyjbrh.com/docs/assistant/source-conflict/)
- [Making a correction actually change behaviour](https://connectbyjbrh.com/research/memory-correction-that-sticks/)

## What this page is based on

- `docs-source/sources/GENERAL.md` §8 — `remember`, `forget` and the tool set
- Connect capability registry (docs-source/facts.py) — `memory_tiers`, `memory_editing`, MEASURED memory budget
- `backend/app/connect_memory.py` — tiers, supersession and directives