TencentDB Agent Memory

TencentDB Agent Memory is Tencent’s memory system for AI agents, delivering fully local long-term and short-term memory via a 4-tier progressive pipeline with zero external API dependencies. Core philosophy: “Agents remember, Humans innovate.”

Overview

The project features a dual memory architecture: Symbolic Short-Term Memory compresses tool logs via Mermaid syntax to reduce token usage and improve task success rates; Layered Long-Term Memory distills fragmented conversations into structured Personas and Scenes. Long-term memory spans four tiers: L0 raw conversations, L1 atomic facts, L2 scenario blocks (Markdown), L3 user profiles — lower layers preserve evidence, upper layers preserve structure.

Benchmark results are impressive: short-term memory achieves +51.52% success improvement and -61.38% token reduction on WideSearch; long-term memory improves PersonaMem from 48% to 76%.

Key Features

• Dual memory architecture: Symbolic short-term memory (Mermaid Canvas) + layered long-term memory (L0-L3 semantic pyramid)
• Fully local: Built on SQLite + sqlite-vec, zero-config, zero external API dependencies
• Hybrid retrieval: BM25 + vector search + RRF fusion ranking
• White-box debuggability: L2 scenarios are plain Markdown, L3 persona in persona.md, readable by humans and agents
• Benchmark validated: Short-term token reduction of 30-61%, long-term accuracy improvement of 59%
• Multi-framework integration: OpenClaw plugin + Hermes Gateway adapter

Use Cases

• Agent long-term memory: Remember user preferences, habits, and interaction history across sessions
• Complex task context compression: Reduce token consumption via Mermaid symbol graphs, improving long-task success rates
• Local privacy protection: Sensitive data never leaves the machine, suitable for finance and healthcare scenarios
• Agent personalization: Achieve per-user customized agent behavior based on user profiles

Technical Highlights

• Language: TypeScript
• Storage backend: SQLite + sqlite-vec (local) / Tencent Cloud Vector Database (TCVDB)
• Node.js requirement: >= 22.16
• OpenClaw requirement: >= 2026.3.13
• Agent tools: tdai_memory_search / tdai_conversation_search
• License: MIT