Agent memory

What agent memory is

Agent memory lets the agent remember things across sessions. Without it, every session starts from zero: the agent doesn't know your preferences, your project context, or what you tried last time. With memory, the agent picks up where you left off.

Memory is a Pro feature. Trial sessions have memory within the session but it doesn't persist after the session ends.

Memory categories

The agent organizes memories into five categories:

| Category | What it stores | Example |

|---|---|---|

| Project | High-level facts about your project | "Training a LoRA adapter for Llama 3 8B on medical Q&A data" |

| Preference | How you like things done | "Always use AdamW optimizer, never use SGD" |

| Finding | Results and conclusions from experiments | "Learning rate 3e-4 causes divergence after epoch 5; 1e-4 is stable" |

| Context | Environment and setup details | "Dataset is in /data/medical_qa_v2, 45K examples" |

| Skill | Learned procedures specific to your workflow | "To evaluate this model, run eval.py with --split test --beam 4" |

How it works

After each session, the agent extracts key memories and stores them. At the start of the next session, it loads relevant memories based on the project you're working on.

You can view and manage memories in Settings > Memory. Each memory shows:

• The content
• The category
• When it was created
• Which session it came from

Delete any memory by clicking the trash icon. The agent won't reference it again.

Memory across sessions

When you start a new session on the same project, the agent loads all memories tagged to that project. If you start a session on a new project, the agent starts fresh (preferences are global, so they still carry over).

This means the agent knows things like:

• What you tried before and what worked
• Your coding style preferences
• Where your data lives
• What evaluation scripts to use

The more sessions you run, the smarter the agent gets about your specific workflow.