Engram Formation System: Bio-inspired Computational Implementation of Memory and Its Applications in Robotics and AI

This article introduces an innovative study: Engram Formation System (EFS), which simulates the engram formation mechanism in the biological brain and applies neuroscience principles of memory encoding and storage to the fields of robotics and AI. Core objective: To provide AI agents with storage and recall capabilities closer to the natural memory of living organisms, supporting multimodal input integration, spatiotemporal pattern recognition, and predictive learning. Original authors: OpenAlex indexed authors Source: openalex Original title: Engram Formation System: Computational Implementation of bioinspired memory for Robotics and AI Research Original link: https://doi.org/10.5281/zenodo.18236561 Publication time: 2026-12-29

Engram refers to the neuronal clusters and their connection patterns that store specific memories, with four key characteristics:

1. Sparsity: Only a small subset of neurons participate in encoding, improving storage efficiency;
2. Distributed nature: Memories are distributed across the connection patterns of neuronal clusters rather than individual neurons;
3. Overlap: Different memories can share neurons, providing a basis for associative learning and generalization;
4. Dynamicity: Engrams are reactivated during recall and may be modified.

Key neural mechanisms of memory formation:

• Synaptic plasticity: Based on Hebbian theory, synaptic connections between synchronously activated neurons are strengthened (LTP/LTD);
• System integration: New memories are first rapidly encoded in the hippocampus, then transferred to the cortex for long-term storage via sleep replay;
• Pattern separation and completion: Distinguishing similar experiences (role of the dentate gyrus) and retrieving complete memories from partial cues;
• Predictive coding: Actively predicting events and using prediction errors to drive learning and memory updates.

The core design of EFS is a dynamic, associative, and predictive memory mechanism, distinct from the static weight storage of traditional AI:

1. Multimodal input processing: Encode visual, auditory, tactile, and other inputs into high-dimensional tensors, map them to a shared representation space, and support cross-modal associative learning;
2. Spatiotemporal pattern detection: Track temporal context through recursive connections and gating mechanisms to identify sequential, conditional dependency, and cyclic structural relationships;
3. Engram encoding and storage:
   • Evaluate the novelty and importance of inputs, and selectively encode salient information;
   • Select a sparse subset from the neuron pool (considering associations with existing memories);
   • Strengthen connections following Hebbian-like rules, adjusted with prediction errors;
4. Predictive memory retrieval: Pre-activate relevant engrams based on the current context, and consolidate or update memories according to the match between predictions and reality.

Three key application scenarios of EFS in robotics:

1. Context-aware navigation: Build experience-based cognitive maps that associate locations with sensory features/events; support predictive navigation (e.g., turning left to reach the charging station);
2. Human-robot interaction learning: Learn user habits (wake-up time, room temperature preferences, etc.), form personalized engrams, and provide customized services; infer whether assistance is needed when detecting behavioral deviations;
3. Tool use and skill learning: Record tool properties, operation sequences, force feedback, and failed attempts to form procedural memory engrams, supporting skill generalization (e.g., transfer of grasping skills).

Core differences between EFS and existing AI memory systems:

• vs Traditional neural networks: Static weight storage (implicit) vs dynamic explicit engrams (independently activatable/modifiable), solving the stability-plasticity dilemma;
• vs Vector databases: Passive similarity retrieval vs active context-predicted activation, supporting associative leaps rather than just similarity matching;
• vs Neural Turing Machines: Content addressing vs context-sensitive prediction, emphasizing tight integration of memory with the perception-action cycle.

Technical challenges and future directions for EFS:

1. Scale and efficiency: To scale to large-scale memory storage, consolidation mechanisms like sleep replay need to be introduced;
2. Memory organization and indexing: Develop hierarchical representations and semantic networks to support efficient retrieval and reasoning;
3. System integration: Design standardized interfaces to seamlessly integrate with existing robot operating systems and AI frameworks;
4. Neuroscience integration: Incorporate new findings (e.g., memory reconsolidation) to optimize memory update mechanisms.

EFS represents an important attempt in bio-inspired AI, creating a more natural and efficient artificial memory system by simulating brain memory mechanisms. Its interdisciplinary value lies in translating neuroscience findings into computational models and demonstrating potential in robotic applications.

In the future, EFS is expected to be applied in scenarios such as personal assistant robots, autonomous driving, educational aids, and medical rehabilitation devices, driving AI evolution from executing preset programs to intelligent agents that can learn from experience, form memories, and guide behavior.