We’re excited to announce the release of ByteRover Beta CLI 0.3.1.

As a developer-first team, our highest priority has always been the same:
improve the quality of memory search and retrieval, because this directly impacts how engineers work with context every day.

ByteRover CLI 0.3.1 is the result of that focus. In this release, we introduce:

• A fully redesigned memory structure - Context Tree

• A brand-new AI retrieval technique - Agentic Search

• Sharper, more reliable search and retrieval across your whole codebase

• Cleaner core commands + a new brv pull command

You’ll notice significant improvements in memory retrieval quality the moment your agent starts retrieving context.

👉 Try the new Beta CLI 0.3.1

👉 Read the documentation

Let us know your feedback when you try it out.

1. A brand-new memory structure: Context Tree

In Beta CLI 0.3.1, memory is now reorganized into a Context Tree - a more well-structured memory system that replaces the old flat organization of previous version. This change dramatically improves how your agents understand, search, and retrieve context.

The Context Tree organizes memory into:

• Domains: High-level categories (e.g., Architecture, API, Frontend).

• Topics: Specific subjects within domains (e.g., Authentication, Components).

• Context Files: Markdown files containing your actual knowledge.

Whenever you store new memories, ByteRover automatically places them into the correct Domain / Topic.

2. From Vector DB → Agentic Search that delivers huge improvement in retrieval quality.

With this new version, we replaced vector DB + similarity search with our new agentic technique "Agentic Search". This enables:

• More accurate, task-specific retrieval.

• Less noise and irrelevant piece of context.

This makes agents retrieve only the exact piece of context you need from each memory, not full chunks of unnecessary details like before.

Throughout building ByteRover, we realize that Vector DB and cosine similarity are not great at capturing complex and high-variance coding details, things that can change quickly. Plus, similarity search is strong at recall but often weaker at precision compared to traditional full-text search. But for code, precision is critical. In programming, retrieving a similar-looking function instead of the correct one can overload your context window with unnecessary details.

And as LLMs are now excellent at tool use and structured data, instead of flattening code into embeddings and hoping cosine similarity finds the right function, we let the agent navigate the codebase like a developer would. That is why we called this new approach Agentic Search.

3- Depreciation of ACE

With Context Tree + Agentic Search in place, we’re officially deprecating the ACE workflow from previous versions.

There are several reasons for this:

• With ACE, rely entirely on coding agents for inputs of memories, if coding agents use our tools inefficiently, we cannot effectively manage context strategies for users. (garbage in, garbage out).

• With ACE and old retrieval tool, coding agents often get a large chunk of memories, only a small part of them is actually needed by coding agents cannot filter out irrelevant parts, which poisons the context of coding agents.

• With ACE, we still use vector db and similarity search for our tools, which is replaced by our "Agentic Search" as shared earlier.

4- Updated Core Commands & "brv pull"

There are a few small changes to the CLI commands:

• brv add → brv curate (store memory into local storage)

• brv retrieve → brv query (retrieve memory).

• And especially, a new command brv pull . This entirely new command helps pull latest update of memories from your team's workspace to your codebase. Just like you git pull latest code from your team's repo.

👉 Try the new Beta CLI 0.3.1

👉 Read the documentation

Looking Ahead

Beta CLI 0.3.1 marks a huge milestone for our team, solving multiple deep technical challenges to bring sharper retrieval, cleaner context, and a significantly better developer experience.

We will be releasing more important features planned for the 0.3.x versions over the upcoming weeks, and we can’t wait to share them with you.

Let us know your feedback when you try it out. You can share your feedback at: support@byterover.dev or hop on a call here.

Best,
ByteRover Team