Zaxy：为AI智能体构建持久化记忆系统的开源框架

Zaxy is an innovative open-source framework developed by syndicalt, focusing on building persistent memory systems for AI agents. It addresses the problem of short-term memory in agents by providing a multi-layered architecture covering event auditing (Eventloom), tamper-proof traceability (hash chain), relational knowledge storage (Neo4j temporal graph), context management (Memory Checkout), and standardized tool integration (MCP). This framework turns agents' workflow into reusable, auditable, and traceable knowledge assets, enhancing their continuous work capability in complex scenarios.

With the rapid development of AI Agent technology, agents can execute increasingly complex task chains. However, a key challenge remains: agents' memory is transient. Context, decision processes, and learning experiences are often lost when tasks end or sessions restart, severely limiting agents' continuous work ability in complex scenarios. Zaxy was born to solve this problem, offering a complete persistent memory system to turn agents' work results into reusable, auditable, and traceable knowledge assets.

Zaxy's core architecture includes:

1. Eventloom: A structured audit log system capturing key events with details like event type, timestamp, participants, input/output, and metadata, enabling complex queries and analysis.
2. Hash Chain Traceability: Using blockchain-like hash chains to ensure event log integrity—each event includes the previous event's hash, making tampering detectable and records verifiable.
3. Neo4j Temporal Graph: A graph database for relational knowledge storage, supporting flexible relationship modeling, efficient association queries, and temporal tracking of entity/relationship evolution.
4. Memory Checkout: Context compression mechanism with relevance retrieval, importance stratification, summary generation, and dynamic loading to fit within LLM context limits.
5. MCP Integration: Standardized Model Context Protocol interface for integration with mainstream agent frameworks (LangChain, AutoGen, Semantic Kernel) via retrieval, capture, and feedback interfaces.

Application Scenarios:

• Customer service agents: Remember customer history for coherent service.
• Code development assistants: Capture project changes and decisions for context-aware suggestions.
• Scientific research assistants: Manage literature notes and concept associations for knowledge discovery.
• Enterprise process automation: Meet compliance requirements with traceable decisions.

Technical Highlights:

• Modular design: Components can be used independently or combined.
• Extensible storage: Supports Neo4j (default) plus Elasticsearch, vector databases.
• Privacy & security: End-to-end encryption, fine-grained access control, data脱敏.
• Performance optimization: Multi-level caching, async indexing, batch processing for high loads.

Zaxy stands out for its comprehensive features compared to other solutions:

Feature	Traditional Logs	Vector Memory	Zaxy
Event Recording	✓	✗	✓
Semantic Retrieval	✗	✓	✓
Relational Graph	✗	✗	✓
Traceability Verification	✗	✗	✓
Temporal Tracking	Limited	✗	✓

This comprehensive architecture makes Zaxy ideal for enterprise-level AI agent memory systems.

Zaxy's future plans include:

1. Federal Memory: Support memory synchronization and sharing among multiple agent instances for distributed memory networks.
2. Active Recall: Proactively identify and prompt relevant historical memory for current tasks.
3. Memory Compression & Distillation: Extract core experiences from massive long-term memory.
4. Cross-modal Memory: Extend support for image, audio, video, and other non-text content storage and retrieval.

Zaxy represents an important direction in AI agent memory systems. It not only solves basic persistent memory problems but also provides practical long-term memory capabilities via multi-layered architecture (audit logs, traceability, knowledge graphs, context management). For teams building production-level AI agent applications, Zaxy is a valuable infrastructure project worth attention and evaluation.