What Is a Replay Buffer in Training? Revisiting Old Examples to Prevent Forgetting
A simple, effective defense against forgetting: keep a sample of old training data around and periodically train on it again alongside new data.
Keeping a sample of old data to revisit later
A replay buffer stores a representative sample of previously seen training data, which then gets mixed back in with new training data during later stages of training — periodically "replaying" a bit of the old material rather than training exclusively on new data going forward.
Why revisiting old data helps, in simple terms
Mixing a sample of old data back into new training gives the model continued exposure to the patterns that data represents, directly counteracting the natural tendency for those older patterns to fade or get overwritten as training focuses increasingly on new material — a straightforward, practical defense against catastrophic forgetting.
The real cost: it isn't free
A replay buffer requires storing and managing that older data, and mixing it in uses training capacity that could otherwise go entirely toward the new task — the size and mixing ratio of a replay buffer is a genuine tuning decision, balancing forgetting prevention against how much capacity is available for new learning.
Frequently Asked Questions
What is a replay buffer in training?
A stored sample of previously seen training data that gets mixed back in with new data during later training, giving the model continued exposure to older patterns rather than training exclusively on new material.
Why does a replay buffer help prevent catastrophic forgetting?
Continued exposure to a sample of older data directly counteracts the natural tendency for those patterns to fade or get overwritten as training increasingly focuses on new material.