The Memories section on an agent's edit page, listing the agent's saved memory files with a name and last-modified date

PushMetrics agents get sharper over time because they've got a place to remember things between conversations: a private notebook called memory.

Every agent has its own. You and the agent both write to it. The next conversation already knows what's in it.

Memory: the agent's own notebook

Every agent has a private memory: a folder of markdown notes it can read whenever it wants and update as it picks things up. Things memory is good for:

  • Workflows the agent should always follow ("the weekly report goes to #analytics first").
  • Numbers and definitions you want it to keep straight ("Q4 starts October 1st", "ARR excludes one-time fees").
  • Templates and stock wording for messages it sends often.
  • Hard rules ("never email customers without my sign-off").
  • Notes the agent writes for itself as it learns the lay of the land.

Adding a memory note

Memory is managed from the agent's edit page.

1
Open the agent in Settings → AI Capabilities → Agents.
2
Scroll to the Memories section.
3
Click + New Memory to write a note from scratch, or Upload .md to import an existing markdown file.
4
Save. The very next conversation will already have it loaded.

How memory keeps the agent on track

The trick with memory is that it's already in the agent's head before the conversation begins. The contents of every memory note get baked into the system prompt the moment someone opens a chat. The agent isn't trying to remember to look something up. The rules are sitting in front of it from message one.

A few things worth knowing:

  • Notes are loaded up front, every time. What you've written is part of the agent's starting context, so it doesn't drop rules just because someone asked a short question.
  • The agent updates its own notes. When it learns something new mid-conversation, it saves it for next time using the same write tools you'd use through the UI.
  • Every change has history. You can see every edit, who made it (you or the agent), and roll back if needed.
  • One notebook per agent. Two agents = two notebooks. What one agent learns doesn't leak to the other.
⏱️
Why "loaded up front" matters. Telling an AI "go check your notes before answering" is unreliable. It might decide it already knows and skip the check. Pre-loading the notes into the system prompt sidesteps that whole problem. The rules are simply present.

A real example

How to send the weekly report
1. Always use the "Weekly Metrics" SQL block for data.
2. Format currency to 2 decimal places.
3. Post the Slack summary to #analytics before sending the email.
4. Email goes to the "Leadership" recipient group.
5. Include a comparison chart (bar, not pie).
6. If any KPI dropped more than 10%, flag it in red.

After saving that note, the agent works through all six rules every time someone asks for the weekly report. You never have to repeat them in chat.

Turning memory on for an agent

Memory is gated by the Memory permission on the agent.

  1. Open the agent under Settings → AI Capabilities → Agents.
  2. Scroll to Agent Permissions → Memory.
  3. Turn on Use Memory.
  4. Hit Save Changes.

That's it. The agent will read its memory notes from the next message onward.

What memory's good for (and what it isn't)

Memory is the right home for things specific to one agent's job: the workflow it always follows, the rules you've set for it, the templates it reuses. The notebook is private to that agent and loaded into context every time.

For facts that need to actually run against the database, like a definition of "active customer" or "monthly recurring revenue", reach for a metrics view instead. Metrics views are queryable. Memory notes are prose.

💡
Memory for things specific to one agent's job. Metrics views for definitions you want the agent (and your reports) to query against the warehouse.