MemoryBear

MemoryBear is a memory management system for AI agents that models the lifecycle of knowledge as perception, extraction, association, and forgetting. Rather than treating stored knowledge as static records to look up, it extracts structured facts from conversations and documents, links them in a graph, and lets unused knowledge decay and expire, drawing on analogies to hippocampal encoding, neocortical consolidation, and synaptic pruning.

The system targets three problems: context loss in single models (window limits, static training data, recency bias), memory silos between cooperating agents, and semantic ambiguity from jargon, colloquial phrasing, and cross-language references. Multiple agents can share one memory store so that user intent and history survive hand-offs.

Architecturally it is a FastAPI service backed by PostgreSQL, Neo4j, Redis, and Elasticsearch, with Celery workers on three queues (memory read/write, document parsing, and periodic reflection) and a React web frontend. LLM access goes through LangChain, OpenAI, DashScope, or AWS Bedrock, and MCP integration uses FastMCP and langchain-mcp-adapters. Two API surfaces are exposed: a JWT-protected management API under /api and an API-key service API under /v1 for extraction, graph operations, search, and forgetting control.

Reported benchmark results compare MemoryBear with Mem0, Zep, and LangMem on F1, BLEU-1, and LLM-as-a-judge scores; the vector (non-graph) variant is reported at 72.90 ± 0.19% overall accuracy and the graph variant at 75.00 ± 0.20%.

Features

• Memory extraction: parses unstructured text into declarative statements, subject-action-object triples, and timestamps, with configurable summaries of 50 to 500 words
• Graph storage: triples sync to Neo4j across 12 relationship types (hierarchical, causal, temporal, logical, and more), with an interactive graph editor
• Hybrid search: Elasticsearch keyword matching combined with BERT-embedding vector retrieval, where semantic search widens candidates and keyword search filters them
• Forgetting engine: each item carries a memory strength updated by usage and association activity; items below threshold pass through dormancy, decay, and clearance
• Self-reflection: a scheduled daily job checks consistency across related knowledge, reassesses value, and re-weights associations
• Service layer: management and service APIs with auto-generated Swagger documentation and Docker deployment
• Async processing: Celery queues sized for memory I/O (100 threads), document parsing (4 prefork workers), and periodic tasks (2 workers)
• Web frontend: React 18 with Ant Design and graph visualisation via AntV X6, ECharts, and D3.js