---
title: "Knowledge & memory"
description: "The durable context agents capture as they work, so future runs start informed instead of relearning your codebase."
---

> Documentation Index
> Fetch the complete documentation index at: https://docs.trevize.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Knowledge & memory

**Knowledge** is the memory that persists across runs. As agents work, they
record **observations** — durable notes about what they decided, fixed, or
discovered — which later runs can search and build on. This is what lets Trevize
get better at *your* codebase over time instead of starting from scratch every
session.

## Observations

Each observation is captured automatically during a [task](/concepts/tasks) or
[session](/concepts/sessions) and has a **type**:

- **Decision** — a choice made and why.
- **Bug fix** — a defect and its resolution.
- **Feature** — something built.
- **Refactor** — a structural change.
- **Discovery** — something learned about the codebase.
- **Change** — a general change.

An observation records a **title**, a **narrative**, structured **facts** and
**concepts**, and the **files** it read and modified.

## Browsing knowledge

The **Knowledge** page summarizes what's been captured — a total plus per-type
counts — and lets you **search** and filter by type. Each observation opens to a
detail view with its narrative, facts, concepts, and file lists. Knowledge is
**read-only insight**: it's captured for you as agents work, not hand-authored.

## How it's used

Observations are semantically searchable (vector embeddings), so when a new run
starts, relevant past work can surface as context. Knowledge is scoped to your
workspace and projects.

> **A workspace-level capability**
>
> Knowledge capture is governed at the [workspace](/concepts/workspaces) level.
> It compounds: the more work runs through Trevize, the richer the context future
> runs begin with.

Source: https://docs.trevize.dev/concepts/knowledge/index.mdx