Overview

TencentDB Agent Memory provides a robust, fully local long-term memory solution for AI agents. It addresses the limitations of traditional memory systems by implementing a sophisticated 4-tier progressive pipeline. This architecture ensures that agents can retain and leverage past experiences without any external API dependencies, leading to more consistent and intelligent agent behavior. The core philosophy is to enable agents to "remember what should be remembered" so that human users can focus on judgment, creation, and tasks that truly matter.

Key Features

TencentDB Agent Memory is built upon two fundamental pillars: memory layering and symbolic memory. This approach moves beyond simple flat storage and irreversible summarization, enabling agents to reason more effectively.

• Layered Memory Architecture: Implements a hierarchical system for both short-term and long-term memory.
  • Short-term context layering: Manages in-task information by archiving raw tool outputs, extracting step-level summaries, and condensing the state into a lightweight Mermaid canvas. Agents attend to the top-layer structure and can drill down to lower layers for details when errors occur.
  • Long-term personalization layering: Builds a semantic pyramid from raw dialogue to atomic facts, scenario blocks, and user profiles (personas). This allows for structured recall of preferences and past experiences across sessions.
  • Skill generation layering: Distills reusable skills and standard operating procedures from execution traces.
• Symbolic Memory with Mermaid: Replaces verbose intermediate logs with high-density Mermaid symbol graphs. This significantly reduces token usage while preserving full traceability through node_id referencing. Full tool logs are offloaded to external files.
• Progressive Disclosure & Heterogeneous Storage: Utilizes a dual-layer storage strategy. Lower layers (facts, logs) are stored in databases for robust retrieval, while upper layers (personas, scenes) are stored as human-readable Markdown files for high information density and inspection.
• Full Traceability and Lossless Recovery: Guarantees a deterministic path from high-level abstractions back to ground-truth evidence, ensuring that information is never lost.
• Zero External API Dependencies: Operates entirely locally, enhancing privacy and reliability.
• Performance Enhancements: Demonstrated to cut token usage by up to 61.38% and improve task success rates by up to 51.52% when integrated with OpenClaw. Long-term memory accuracy, such as for PersonaMem, can be increased from 48% to 76%.

Typical Use Cases

• AI Agents for Complex Workflows: Enables agents to maintain context and learn user preferences over extended periods, crucial for tasks like software development, research, or intricate problem-solving.
• Customer Support Bots: Allows bots to remember past interactions, customer history, and specific solutions, leading to more personalized and efficient support.
• Personalized Assistants: Agents can build a deep understanding of individual user habits, preferences, and ongoing projects, offering proactive and context-aware assistance.
• Long-Horizon Task Execution: Essential for agents operating on tasks that span numerous steps or interactions, where retaining fine-grained details and overall progress is critical.
• Reducing LLM Token Costs: By intelligently compressing and managing conversational history and tool logs, it significantly lowers operational costs associated with large language models.