Reliable perception
Distinguish an object from its background, view and pose. Keep unresolved evidence separate from confirmed identity.
TECHNOLOGY / LIFE CORE
Experience becomes
an operating advantage.
A research architecture for perception, predictive memory and continual learning within explicit resource limits.
THE ARCHITECTURE
Salience-Gated
Predictive Memory.
Life Core investigates how an intelligent system can retain what matters, anticipate new observations and continue learning from where it left off.
Salience-Gated Predictive Memory (SGPM) is the descriptive name of our architecture. It brings together immediate interpretation, selective retention and slower consolidation. The central research question is whether acquired experience can improve later perception and behaviour under the same memory and computing budget.
Our inspiration comes from the distinction between rapid experience and longer-term learning. Biological inspiration guides questions; it does not establish biological fidelity or a new scientific mechanism.
A cycle of evidence.
Our long-term architecture connects sensing, expectation, action and learning. The physical action loop remains a development objective.
- 01
Observe
Acquire a versioned observation and its source.
- 02
Anticipate
Form an expectation before the outcome arrives.
- 03
Compare
Measure the difference against fresh evidence.
- 04
Learn
Update a bounded model and retain useful experience.
- 05
Consolidate
Use quiet replay to test longer-lasting learning.
PERSISTENT ACQUIRED STATE
A life file.
A continuing learner.
A life file preserves acquired state so that learning can continue across sessions.
The research goal is to retain the model state, memory, calibration and continuation information needed for the same next update after saving and restoring a system. This makes continuity testable.
A small saved file is one part of the resource picture. We also count perception models, working memory, replay, processing time and prior training. Existing prototype readers include disclosed pretrained components; the longer-term ambition is control over the complete learning stack.
Our current experiments preserve learned expectations after temporary replay records are removed. That is a narrow form of consolidation, with explicit limits on what was learned. It is not a stored human life or complete robot skill library.
Read the current evidence ↗What has to work together.
The development agenda spans the whole learning loop, with capability claims tied to individual tests.
Useful memory
Show that retained experience improves later performance against a strong comparison system with the same resources.
Action and consequence
Predict what movement will change, then measure the actual outcome. A commanded motion is not evidence of successful action.
Continual adaptation
Acquire new capabilities while testing retention of earlier ones. Preserve the full state needed to resume learning.
BUILDING THE NEXT CHAPTER