# Part 6: The read path | Team0

> Give agents useful, current, inspectable context from the memory format you choose.

Agent-readable version of https://team0.ai/learn/the-read-path. 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 6

# The read path

Give agents useful, current, inspectable context from the memory format you choose.

Yogev Ben-Tov · 3 min read

A brain can hold the right information and still give an agent the wrong context. The question is how to supply enough relevant, current information for the task, within a budget you can operate.

By the end of this part, you will have a tested way for an agent to read your prototype, whether its memory is MD files, database records or another system.

## Define what a useful read contains

Choose a few real tasks: a question about a person, preparation for a meeting, a decision about a topic and a follow-up to an earlier question.

For each, list the information the agent needs, its sources and any uncertainty that changes the answer. Include an unknown person and two people with the same name. No matching record, ambiguous identity and a failed read need different responses.

## Choose retrieval for your scale

Start with what your storage already supports. A small collection of MD files may work with an index and text search. Larger collections may benefit from keyword, semantic or relationship-based retrieval. Names, numbers and exact codes are useful cases to check before adding a new search method.

Ask your agent to implement one improvement where the test shows a miss. Compare it with the existing read rather than assuming a more elaborate retrieval system is better.

Enforce the caller's access before returning context. A model instruction to keep a retrieved secret private is not an access boundary.

## Keep the answer inspectable

An agent should be able to tell which information is current, which is uncertain and where it came from. Historical values should remain distinguishable from today's plan. Return enough to support the task without flooding the agent with unrelated detail.

Choose a response-size and latency budget. A timeout or unavailable source should be visible. Source freshness matters: an answer based on last week's copy may need a caveat or a live check.

For a prototype, one `read(question)` interface can keep experiments manageable. An existing file-reading workflow is fine too. The interface is a choice; the answer quality is what you measure.

## Measure

```
python3 score.py questions.jsonl part5.jsonl part6.jsonl --out part6-scores.json
```

Check relevance, ordering and omissions together. Fewer tool calls may help, but compare latency, token use and costs as well. One large read can cost more than several small ones.

Next, part 7: making agents use it.

---

*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).*