SuanmoSuanyangTechnology/MemoryBear

What it solves

MemoryBear addresses the limitations of traditional AI memory, such as the restricted context windows of LLMs, the static nature of training data, and the "recency bias" where AI forgets early conversation history. It also solves data silos in multi-agent systems where different agents cannot share user preferences or history, and semantic ambiguity in reasoning caused by jargon or context-dependent references.

How it works

MemoryBear implements a cognitive-inspired memory lifecycle: perception, extraction, association, and forgetting. It uses a Memory Extraction Engine to parse unstructured data into structured triples (subject-action-object) and temporal tags. These are stored in a Neo4j graph database to model complex relationships. To retrieve information, it employs a Hybrid Search mechanism combining keyword matching (Elasticsearch) and semantic vector search (BERT). Finally, a Memory Forgetting Engine mimics biological synaptic pruning by decaying the strength of unused knowledge over time, while a Self-Reflection Engine periodically checks for logical consistency and optimizes association weights.

Who it’s for

It is designed for developers building advanced AI agents and multi-agent systems that require long-term, evolving, and personalized memory across different sessions and agents.

Highlights

• Graph-First Architecture: Uses Neo4j to support millions of entities and complex relational edges.
• Synaptic Pruning: A dynamic forgetting mechanism that removes redundant knowledge to reduce waste.
• Hybrid Retrieval: Combines vector and keyword search for 92% retrieval accuracy.
• Self-Reflection: Automated daily processes to detect logical conflicts and reinforce high-value knowledge.
• High Performance: FastAPI service layer sustaining 1000 QPS with sub-50ms latency.