AmeforgeAmeforge
R&D · Active ResearchInternal name: LPOLMemoryPrototype Functional

MemoryBank

Persistent Domain-Segmented Memory for Continuous Learning

MemoryBank is a persistent external memory system that enables neural networks to accumulate knowledge across multiple learning sessions without forgetting previously acquired information. The memory is organized by domain, allowing efficient retrieval through similarity-based lookup.

Preliminary results only. Validation is ongoing. Not yet available for commercial use. Demonstration available on request for qualified partners.

Catastrophic forgetting limits continuous learning

When a neural network learns a new task, it tends to overwrite the weights encoding previous knowledge. This phenomenon, known as catastrophic forgetting, forces practitioners to retrain models from scratch whenever new data or domains are introduced.

Existing solutions such as Elastic Weight Consolidation or Progressive Neural Networks either impose significant computational overhead or require architectural modifications that limit scalability.

External memory segmented by domain

Rather than modifying the model weights to accommodate new knowledge, MemoryBank stores information in an external memory organized by domain. When the model encounters a new input, it retrieves relevant memories through similarity-based lookup before generating a response.

Key property: Because knowledge is stored externally rather than encoded in weights, adding new domains does not interfere with existing ones. The model weights remain stable across learning sessions.

Four components, one persistent memory

DC

Domain Classifier

Automatically identifies the domain of incoming data (code, medical, legal, general) before routing to the appropriate memory segment.

MS

Memory Slots

4096 addressable memory slots organized by domain. Each slot stores an encoded representation of a knowledge unit.

SR

Similarity Retrieval

Cosine similarity-based lookup retrieves the most relevant memories for each input query, enabling context-aware generation.

PL

Persistence Layer

Memory state is preserved between sessions, allowing the system to accumulate knowledge over time without retraining.

Memory Slots
4 096
Retrieval Method
Cosine Similarity
Domain Classifier
Integrated
Persistence
Cross-session

10 sequential learning cycles, 3 domains

Results are preliminary and based on tests conducted on datasets of moderate size. Large-scale validation is ongoing. These figures should not be interpreted as production-ready benchmarks.

0
Forgetting Events
Over 10 sequential learning cycles
3
Domains Tested
General, Code, Medical (public datasets)
10
Learning Cycles
Consecutive sessions without degradation

Limitations of current testing: Tests were conducted on datasets of moderate size only. Performance on complex tasks and large-scale validation is currently under evaluation. No comparison with published baselines has been conducted yet. Results are reported as measured, without extrapolation.

Systems that learn over time

Conversational assistants
Accumulate user-specific knowledge across sessions without retraining the base model
RAG systems
Augment retrieval-augmented generation with persistent long-term memory
Continuous learning pipelines
Integrate new data incrementally without full retraining cycles
Domain adaptation
Specialize a general model for a new domain by adding domain memory, not retraining weights

Active research, not yet commercialized

Implemented and functional

Core memory system is operational in laboratory conditions. Similarity retrieval and persistence layer are validated.

Ongoing validation

Performance on complex tasks and large-scale datasets is currently being evaluated. No publication date set.

Access

Demonstration available on request for qualified research partners and investors. Contact: contact@ameforge.tech