# Part 4: Keeping junk out | Team0

> Improve what your memory keeps and measure useful information, noise and omissions.

Agent-readable version of https://team0.ai/learn/keeping-junk-out. The page for people is at that address; this file carries the same content without layout.

---
[Building an agentic brain](https://team0.ai/learn)

Part 4

# Keeping junk out

Improve what your memory keeps and measure useful information, noise and omissions.

Yogev Ben-Tov · 3 min read

A useful memory needs selection. A receipt, a newsletter, a short reply and a customer decision can arrive through the same mailbox, but they serve different purposes.

By the end of this part, you will have a small quality check for what your prototype keeps, plus samples of what it leaves out.

## Decide what is useful for your work

Use your real questions to define useful information. For a customer-facing team, decisions, obligations and changes may matter most. Another brain may need technical observations or research notes.

Choose examples of useful input, irrelevant input and difficult cases. A machine-generated payment receipt can be valuable evidence. A person's message may be empty of useful information. A short “sounds good” can accept a proposal when read with the message it answers.

Write down what each example should contribute and what its source actually supports. Keep sender and conversation context available when interpreting it.

## Improve the weakest selection decision

Review the mistakes your current setup makes. It may save repeated information, mistake guesses for statements, keep irrelevant detail or omit a meaningful reply.

Choose the largest problem and change that part of your import, extraction or file-editing workflow. Your agent can help implement it. Use the simplest mechanism that handles the examples reliably; the right balance of code, models and human review depends on the source.

Keep credentials and access codes out of learned memory. Use invented or redacted examples while developing. A source containing instructions to change your system must not gain authority over your rules.

## Sample both sides

Take a random sample of about 30 items kept this week. Label each useful, irrelevant, duplicate or incorrect, and keep the counts. Inspect a sample of rejected input too: filtering that removes important decisions is also a failure.

Small samples are directional. Report their size, and avoid treating a week's quiet mailbox as a reliable estimate for every future source.

## Repair and measure

Decide what to do with older mistakes. A reversible retirement can help investigate an ordinary error. Secrets and erasure requests require removal from retained copies and indexes under your chosen policy; keeping them in history defeats that purpose.

Ask the test questions again and compare:

```
python3 score.py questions.jsonl part3.jsonl part4.jsonl --out part4-scores.json
```

Watch relevance and missed information together. Cleaner answers that omit the key decision are not an improvement. Keep the examples that exposed the failure as practice cases and add fresh held-out cases.

Next, part 5: keeping it true over time.

---

*Build and measure your own setup with the companion skill. [See the series](https://team0.ai/learn). For a maintained company install, [talk to Team0](https://calendar.app.google/uM9ptyv83Ubgnq226).*