SuanmoSuanyangTechnology/MemoryBear

What it solves

MemoryBear addresses the limitations of traditional AI memory, such as the small context windows of LLMs, the static nature of training data, and the "data silos" that occur when multiple AI agents cannot share information. It aims to move AI from passive data retrieval to proactive cognitive assistance by emulating biological memory processes.

How it works

MemoryBear implements a lifecycle of perception, extraction, association, and forgetting:

• Extraction: It parses unstructured conversations and documents into structured triples (subject-action-object) and temporal tags.
• Storage: It uses a graph-first architecture powered by Neo4j to model complex relationships (causal, temporal, logical) and PostgreSQL for primary data.
• Retrieval: A hybrid search engine combines Elasticsearch for exact keyword matching and BERT embeddings for semantic vector search.
• Forgetting: A "synaptic pruning" engine dynamically decays knowledge based on usage frequency and time, removing redundant data to maintain efficiency.
• Reflection: A scheduled process checks for logical conflicts and optimizes relationship weights based on recent usage.

Who it’s for

Developers building complex AI agents or multi-agent systems that require long-term, evolving memory, personalized user history, and the ability to maintain consistent state across different specialized agents.

Highlights

• Biological Inspiration: Mimics the hippocampus and neocortex for encoding and consolidation.
• Hybrid Search: Combines keyword and vector retrieval for 92% accuracy.
• Dynamic Forgetting: Reduces waste by over 60% by pruning low-value knowledge.
• Graph-Based Memory: Supports millions of entities and tens of millions of relational edges via Neo4j.
• Self-Reflection: Automatically detects logical conflicts and reinforces high-value information.