Agent Memory and Context Window

Memory stores selected information across threads, while the context window limits how much information can be taken into account within a thread.

The awork Agent uses conversation context and personal memories. Find out what carries over between chats and how to get better results.

What the awork Agent remembers

Within a thread, the Agent takes into account the conversation so far, along with any existing artifacts, meaning files and results shared or created in the thread, and the associated context. After context compression, only a shortened summary of this information may be available.

Across different chats, awork currently supports personal memories for the awork Agent. These can be used to save lasting preferences or facts when you explicitly ask.

Personal memories are retrieved once per thread: the first successful retrieval creates a snapshot that is reused for all subsequent messages in the same thread. Even if no matching memories are found during retrieval, it still counts as a successful retrieval. This means a new search is not triggered for every message. If retrieval fails or is disabled, no snapshot is saved, so a later run can try again.

What the Agent does not remember across threads

Not everything from a chat automatically becomes lasting knowledge for the Agent.

• No automatic learning from chats: Neither the Agent nor the model learns from your chats.

• No automatic memory for custom Agents: Custom specialized Agents do not automatically carry information from one thread over to new threads. If you want an instruction to be taken into account in the future, you can explicitly ask the Agent to update its prompt for future threads, or add the new instruction manually.

• Separate memories in shared threads: The memory state in a thread belongs to the user who created it. Other participants in the same thread do not get access to that user's personal memories and cannot take over or replace the saved state.

How to get better results

Good context makes a big difference. It's especially important to keep tasks clearly separated.

Start a new thread for new tasks

Each chat thread has a limited context window. The longer a thread gets, the more older content needs to be compressed. Details can be lost in the process.

That's why we recommend starting a new chat for new, focused, or unrelated tasks. Short, clearly defined threads work better than long threads with changing or mixed tasks.

Use Docs for recurring context

For recurring tasks, you can add context to projects or Spaces. This can include files, instructions, and background information.

That way, you don't have to include this context again in every new thread.

Be explicit about what should remain relevant

If you want the awork Agent to take something about you into account over time, you can say so explicitly.

For example, when working on a task, you can mention that a particular preference should also be taken into account in the future.

Provide the right context

Mention relevant tasks, projects, or documents, or add them with @. You can also start the Agent directly from the relevant task or project.

This gives the Agent the right context for the work at hand.

FAQ

Does the awork Agent learn from my chats?

No. Neither the Agent nor the model learns from your chats. However, the awork Agent can save personal memories and take them into account in future threads. For custom specialized Agents, you can explicitly update the prompt for future threads.

Why should I start a new thread for new tasks?

The longer a thread gets, the more older content needs to be compressed because the context window is limited. Details can be lost in the process. New, focused threads keep the context clear and reduce the risk of older information being lost during context compression.