Sparse Agent Memory gives agents cloud memory they can use across sessions. Working preferences, decisions and useful solutions can be saved and found again when a later task needs them. You can also query, check and back up your memory on the web.
What is agent memory?
When you work with an agent, some information keeps coming up: how you want reports written, which project conventions matter, and how you solved a recurring problem. Keeping that information in memory reduces the need to explain it again in a new session.
For example, you might ask an agent to start daily reports with results, then list unfinished work and next steps. An agent connected to memory can look up that preference when it needs to write another report.
Memory belongs to your account and can be managed by project. Different accounts have separate memory. In the Sparse workspace, each agent uses its own memory; connecting another agent to the same service does not automatically give it access to that content.
Using memory during a task
After connecting a memory tool or enabling Sparse cloud memory, you can use it at three points in your work. The prompts below are examples.
- Recall relevant context before starting. Ask the agent to find conventions and lessons related to the task: “Before you start, check this project's testing and daily reporting requirements.”
- Save information worth reusing. When you correct a working habit or find a useful solution, say: “Remember the approach we confirmed here so you can refer to it next time.”
- Use it in a later task. In a new session, ask an agent connected to memory to look up the current question and apply what it finds alongside the latest information.
Whether an agent reads or updates memory depends on its integration and task instructions. You can also query it yourself on the web, describing the question in ordinary language, such as “How did we solve this problem last time?”
What is useful to remember?
| Information | Example |
|---|---|
| Working preferences | Start reports with results, then unfinished work and next steps. |
| Project conventions | Which checks to complete after changing a particular feature. |
| Confirmed decisions | Which approach was chosen and why. |
| Reusable methods | How to complete a setup or handle a recurring problem. |
| Corrections and lessons | What went wrong last time and how to avoid it. |
Prefer information that will affect future work. Mark temporary figures and short-lived status with their applicable dates, and update them when circumstances change. Keep passwords and keys in a dedicated credential manager.
Getting started in the Sparse workspace
Sign in to the Sparse Agent workspace, open Settings → Cloud memory, and choose Connect Chat account. Authorize with the Auto-Coder Chat account associated with your workspace. You usually do not need to copy an API key.
Check the connection status and memory query model on that page. The automatic option keeps an existing cloud query configuration first and can reuse your Agent model when no configuration is available. You can also choose to keep a separate cloud model.
The setting applies to existing and new agents at a later task activation. Agents with existing local memory can use Migrate existing memory on the same page and show the migration progress. Their local records are kept.
Start with a preference you can easily verify. Ask a connected agent to remember it, then look it up in a later session:
Remember this: when writing my project daily report, start with completed work, then list unfinished items and next steps.
Look up my project daily reporting preferences, then use them to write today's report.
Check both the lookup result and the final report to confirm the preference was used. If the page asks you to repair authorization or configure a model, complete the connection check first.
View, update and manage memory
Open Sparse Agent Memory and sign in with Auto-Coder Chat. In My projects, select your project to query memory, inspect its status, run validation or export a backup. Existing members can sign in directly; first-time members should follow the invitation prompts on the page.
When a convention changes, ask a connected agent to update the relevant memory and describe the new conditions. For example: “Send daily reports in the morning from now on. Update the previous timing preference.” Query again to check that the current requirement can be found.
The web interface also supports importing content to restore or replace a project. Import replaces the project's current content, so export a backup first. To stop using cloud memory, choose Disconnect under Settings → Cloud memory in the Sparse workspace. Agent access stops, and previously saved cloud memory is kept.
Example: ongoing project work
Suppose an agent checks a project every day and reports its progress. During the first task, agree on the reporting requirements and checks, then ask it to save those conventions.
A few days later, a recurring issue appears. The agent can look up the earlier solution and check whether it fits the current situation. If the old approach needs changing, save the updated method together with the conditions where it applies. A later task can then refer to that lesson.
This is an illustrative workflow. Assess it with specific questions: can the saved preference be recalled, does the report apply it, and can the current requirement still be found after an update?
How memory and guides work together
A guide explains how to complete a task. Memory adds your preferences, project decisions and previous experience. An agent can look up the relevant memory before following the guide, so it knows which requirements apply to this task.
Start with one preference and one solved problem to see whether this workflow fits your work. To connect your own agent or program, continue to the memory service usage documentation. In the Sparse workspace, connect directly through Settings → Cloud memory.
Further reading: the reference post and its linked Devin Memory and Dreaming guide.