MemoryBear

MemoryBear is a memory system for AI agents from RedBear AI that treats knowledge as something that changes over time rather than static data to retrieve. Modeled on how the brain encodes, consolidates, and prunes memories, it runs knowledge through a lifecycle of perception, extraction, association, and forgetting. It targets three problems: context lost to model window limits and recency bias, isolated memories across collaborating agents, and semantic drift from jargon, colloquial language, and multilingual input.

An extraction engine parses conversations and documents into core statements, entity-relation triples, timestamps, and summaries. Triples sync to a Neo4j knowledge graph with 12 core relationship types, and hybrid search pairs Elasticsearch keyword matching with BERT embedding search. A forgetting engine gives each item a memory strength that rises with use and decays over time, moving weak items through dormancy, decay, and clearance, and a daily self-reflection job checks for conflicting records, reweights associations, and flags suspicious entries for human review.

It runs as a self-hosted FastAPI service with a management API and a service API for extraction, graph operations, search, and forgetting control, backed by PostgreSQL, Redis, and Celery worker queues, plus a web frontend. Model access goes through LangChain with OpenAI, DashScope, or AWS Bedrock, and MCP support comes through fastmcp. Published benchmarks compare it with Mem0, Zep, and LangMem on F1, BLEU-1, and LLM-as-a-judge scores, reporting 72.90% overall accuracy for the vector version and 75.00% for the graph version.