New data. The same memory budget.
As a model keeps learning, new examples arrive. Only a limited selection of earlier examples can stay in its replay buffer.
MEMORYSAFE LABS / CONTINUAL-LEARNING RESEARCH
Learning continues. What stays?
A model cannot revisit everything it has ever seen. We study how to choose what it returns to.
MVI · PROTECTSCORE · REPLAY
What gets another chance?
Inside a learning memory
Notice what may be fragile.
MVI tracks a smoothed signal from each example’s training loss. It is a signal of vulnerability that helps identify examples needing another look.
Give the selection a reason.
ProtectScore combines MVI with configured value, rarity, criticality and task-age signals. It guides retention and replay; separate quotas set minimum retention for selected groups.
Change what gets practiced.
Within the same memory limit, selected earlier examples return to training alongside new data. Their replay losses refresh MVI, feeding the next decision.
New examples
Fixed memory budget⌑ Separate retention floor
EarlierMVI
Rare
Recent
Earlier
Recent
RecentGNew
Earlier knowledge, limited spaceVulnerability becomes visiblePriority + retention constraintsSelected examples get another look
Learning model
Learning continuesLearning continuesLearning continuesNew data + replay
Replay loss refreshes MVI
The audit trail
EXAMPLE ENTRIES- ADMITA staysA free slot was available.
- REJECTH stays outIts ProtectScore is below the eligible resident’s.
- EVICTF → GG outranks F within their group. B’s group retains its floor.
Run records can include the rule, scores, training step and buffer state, so the selection can be inspected.
RESEARCH SETTING · POSSIBLE APPLICATIONS
Where forgetting has consequences.
Our continual-learning evaluation uses medical images. Financial models, edge devices, and robotics are possible applications to explore with partners.
Medical AI
Rare clinical cases can be overwritten as models absorb new data.
Prioritize vulnerable, clinically important samples.
Model-risk & retraining
Changing data can make previously reliable financial models drift.
Preserve important evidence across retraining cycles.
Edge AI
Strict storage and compute limits make every retained sample matter.
Allocate bounded memory intentionally.
Robotics
New environments can interfere with previously learned skills.
Protect fragile capabilities during adaptation.
This is research into model training. It is separate from the free local assistant plugin and the hosted edition in development. No performance results are presented on this page.