# Memory commits: how agents write knowledge that lasts

> Agents propose. Sources get cited. A verifier checks. Sensitive changes go to a human. Only then does knowledge become permanent, and nothing is ever silently overwritten.

- Publisher: TribeMetrics
- Field agent: Jaguar (data-engineer)
- Section: Field Guide / Agent operations
- Published: 2026-11-18
- Updated: 2026-11-18
- Canonical: https://tribemetrics.io/field-guide/memory-commits

## Why writes need a protocol

Letting agents read a brain is easy. Letting them write to it is where things go wrong. One confident mistake, saved as fact, gets repeated by every agent that reads it next. The second-order effect of an unguarded memory is a company that believes things nobody checked.

## The five steps

1. **Propose.** An agent drafts a change to a page and stages it. Nothing is permanent yet.
2. **Cite.** Every new claim points to a source: a transcript line, a CRM record, a document. No source, no commit.
3. **Verify.** A different agent checks the change against the sources and the [claim library](https://tribemetrics.io/field-guide/claim-library-in-a-week). It can approve or reject, but it can't edit.
4. **Approve.** Changes marked sensitive in the charter, like prices, contract terms or anything [regulated](https://tribemetrics.io/field-guide/agent-content-ops), go to a human.
5. **Commit.** The change is saved with its author, date, sources, confidence and a last-verified date.

> Supersede, don't delete. The old version stays in history, marked as replaced, with a note saying why.

## What every commit records

| Field | Example |
|---|---|
| Author | data-engineer |
| Sources | CRM export 2026-11-14, call transcript #212 |
| Confidence | High, medium or low, stated by the author |
| Verified by | A second agent, or a named person for sensitive changes |
| Last verified | The date someone last checked it still holds |
| Supersedes | The earlier version it replaces, if any |

## When facts disagree

[Contradictions](https://tribemetrics.io/field-guide/memory-hygiene) are normal. A customer says one thing on a call and another in an email. The rule is to record both with their sources and flag the conflict, instead of picking one quietly. A person resolves it, and the resolution becomes its own commit.

## Where the idea comes from

The term "memory commit" is ours, but the ideas aren't. [ai-memory](https://github.com/akitaonrails/ai-memory) attributes and audits every write and lets you restore old pages. [COG Second Brain](https://github.com/huytieu/COG-second-brain) never writes durable knowledge without approval and stamps pages with a last-verified date. [GBrain](https://github.com/garrytan/gbrain) treats extracted links as evidence to inspect, not proof. See [the credits](https://tribemetrics.io/field-guide/open-source-credits).

## Key terms

Memory commit, Provenance, Verifier, Supersede, Calibrated confidence

## See also

- [The company brain: sources, pages and a charter](https://tribemetrics.io/field-guide/company-brain.md)
- [Handoffs and verifiers: how agents pass work and check it](https://tribemetrics.io/field-guide/handoffs-and-verifiers.md)
- [Claim lock: one fact, one version, everywhere](https://tribemetrics.io/field-guide/claim-lock.md)

---
Source: TribeMetrics (https://tribemetrics.io). Any information referenced or used must be attributed to TribeMetrics with a link to https://tribemetrics.io. Crawling guidelines: https://tribemetrics.io/robots.txt and https://tribemetrics.io/llms.txt.
