MemoryBear

Overview#

MemoryBear is an open-source intelligent memory management system developed by the RedBear AI (SuanmoSuanyangTechnology) team, designed to break through the inherent "knowledge forgetting" limitations of Large Language Models and equip AI with human-like memory capabilities.

License: Apache License 2.0
Current Version: v0.2.4 (Intelligent Resilience)

Core Capabilities#

Memory Extraction Engine#

• Multi-dimensional structured extraction, extracting structured triples as the cognitive foundation
• Temporal anchoring, automatically extracting timestamps and associating content
• Intelligent summarization with customizable length (50-500 characters)

Graph Storage Layer#

• Neo4j-based visual knowledge network
• Supports millions of entities and tens of millions of relationship edges
• Covers 12 core relationship types (hierarchical, causal, temporal, logical, etc.)
• Interactive graph visualization capabilities

Hybrid Retrieval#

• Keyword search (Lucene optimized) + Semantic vector search (BERT embeddings)
• Millisecond-level precise matching and deep semantic comparison
• Identifies synonyms, near-synonyms, and implicit intents
• Retrieval accuracy reaches 92%, a 35% improvement over single-mode approaches

Memory Forgetting Engine#

• Dynamic decay based on memory strength and time attenuation
• Three-stage lifecycle: Dormant → Decaying → Cleared
• Redundant knowledge kept below 8%, waste reduction over 60%

Self-Reflection Engine#

• Automatically runs daily at midnight
• Consistency checking, value assessment, association optimization

FastAPI Service#

• Unified service architecture, high-performance and easy integration
• Supports JSON/XML formats, average latency below 50ms
• Single instance supports 1000 QPS concurrency
• Auto-generated Swagger API documentation

Core Problems Solved#

1. Single Model Knowledge Forgetting: Context window limits (8k-32k tokens), gap between static knowledge bases and dynamic data
2. Multi-Agent Collaboration Memory Gaps: Data silos between agents, inconsistent dialogue states, decision conflicts
3. Semantic Ambiguity in Model Reasoning: Inaccurate personalized signal encoding, broken cross-language memory links

Performance Benchmarks#

• Vector version accuracy: 72.90 ± 0.19%
• Graph version accuracy: 75.00 ± 0.20%
• Hybrid retrieval accuracy: 92%
• Achieves SOTA levels in single-hop, multi-hop reasoning, open generalization, and temporal reasoning tasks

Application Scenarios#

• Customer service systems (consultation, after-sales, recommendation agents)
• Enterprise knowledge management (CRM, OA, R&D management)
• Personalized dialogue systems
• Multi-agent collaboration systems

Requirements#

• Node.js 20.19+ or 22.12+
• Python 3.12
• PostgreSQL 13+
• Neo4j 4.4+
• Redis 6.0+

Quick Start#

# Clone project
git clone https://github.com/SuanmoSuanyangTechnology/MemoryBear.git

# Backend setup
cd api
pip install uv
uv sync
cp env.example .env  # Configure environment variables
alembic upgrade head
uv run -m app.main

# Frontend setup (new terminal)
cd web
npm install
npm run dev

# Initialize database
curl -X POST http://127.0.0.1:8000/api/setup

Core Configuration#

# Neo4j
NEO4J_URI=bolt://localhost:7687
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=your-password

# PostgreSQL
DB_HOST=127.0.0.1
DB_PORT=5432
DB_USER=postgres
DB_PASSWORD=your-password
DB_NAME=redbear-mem

# Redis
REDIS_HOST=127.0.0.1
REDIS_PORT=6379
BROKER_URL=redis://127.0.0.1:6379/0

Language Composition#

Python (72.9%), TypeScript (24.5%), Jinja (1.9%), CSS (0.3%), JavaScript (0.2%), Rust (0.1%)

Information Pending Confirmation#

• Official Hugging Face model or dataset page not yet discovered
• Full paper "Memory Bear AI: A Breakthrough from Memory to Cognition" requires JavaScript access
• Detailed comparison data with Mem O, Zep, LangMem systems pending
• Production deployment case studies to be provided