How we built memory review flow

At Augment, context is our moat. Agents are only as useful as the context they can keep track of, and memory is the backbone of that context.

But memory is also fragile. If irrelevant or low-quality details sneak in, users lose trust. If memories are hidden away, they feel out of control.

That’s why we built Memory Review Flow — a way to see, edit, and curate memories as they’re created, so your agent remembers what matters and forgets what doesn’t.

Why we built it

Before Memory Review Flow:

• Agents automatically generated memories.
• Users had no visibility into what was being stored.
• The only way to audit was periodically opening the raw memory file.

This led to unnecessary or low-quality memories piling up, and a frustrating lack of control.

We heard the feedback — from internal teams and from users — and decided to make memory reviewable, editable, and transparent.

How we decide to create a memory

The agent decides to create a memory when it sees something worth persisting across sessions.

Typical examples:

• A long-term project goal mentioned in chat.
• A decision made during debugging or planning.
• A relevant piece of code or system detail.

But not every piece of context deserves to last forever. That’s why we built the review step — so you decide what sticks.

How we use memories in the agent loop

Here’s how the flow works today:

Conversation
   ↓
Agent proposes memory (draft)
   ↓
Memory appears in Turn Summary ("1 Pending Memory")
   ↓
User clicks → review screen opens inside Chat
   ↓
User options:
   - Approve (add to workspace long-term memory)
   - Edit (curate before saving)
   - Discard (reject entirely)
   ↓
Agent loop continues with curated memory context

By keeping this loop lightweight, users stay in the flow of conversation while still curating memory quality.

The technical challenge

The main challenge wasn’t how to create memories — it was how to surface them without breaking UX.

Our solution:

• A new modal directly in the chat panel.
• Inline review tools (approve, edit, discard).
• Turn summary entry (“X Pending Memory”) that acts as the trigger.

This design makes memory review part of the natural chat loop, instead of a separate audit process.

Use cases

Memory Review Flow is especially useful for:

• Opinionated users who want to curate memories for accuracy.
• Teams on long-running projects, ensuring only relevant context carries forward.
• Catching spurious entries early, instead of sifting through a file later.

Looking ahead

Memory Review Flow is just the first step. By showing how we decide to create memories, and letting users intervene before they’re finalized, we’ve closed the gap between automation and trust.

This is how agents get better: not just by remembering, but by remembering well.