TencentDB-Agent-Memory

Product Introduction

1. Definition: TencentDB Agent Memory is a centralized, team-level memory system for AI agents. It is a technical solution categorized as an AI agent orchestration and knowledge management platform. It functions as a persistent memory hub that transforms unstructured data into structured, reusable assets.
2. Core Value Proposition: It exists to solve the critical problem of fragmented and non-reusable knowledge in AI agent teams. Its primary value is enabling consistent knowledge sharing and improved agent capabilities across different frameworks, preventing agents from starting from scratch in every new session or task. This directly addresses the inefficiency and high computational cost of repetitive agent learning.

Main Features

1. Chat Memory (Layered Memory Architecture): This feature automatically distills raw conversation logs (L0) into structured, searchable memory assets across multiple layers (L1 Atoms, L2 Scenarios, L3 Persona). It works by using an asynchronous processing pipeline that extracts facts, preferences, and contextual blocks. Retrieval employs a hybrid search strategy combining BM25 for keyword matching, vector embeddings for semantic similarity, and Reciprocal Rank Fusion (RRF) to merge results, all bounded by token and timeout limits to manage context window usage efficiently.
2. Skill Library: This feature allows agents to extract and manage reusable workflows, or "Skills," from successful task completions and tool call histories. A Skill is not just a prompt but a versioned asset with defined execution steps, trigger conditions, validation rules, and associated resource files. Skills can be privately held, shared within a team, or explicitly assigned to specific agents via the Memory Hub control panel, enabling expertise transfer.
3. LLM-Wiki & Code-Graph (Structured Knowledge Assets): These are two distinct but complementary knowledge asset types. The LLM-Wiki automatically converts documents (product specs, runbooks) into interlinked, structured pages inspired by knowledge base systems. The Code-Graph feature indexes entire code repositories, parsing symbols, files, and call relationships to build a semantic graph. Agents can then perform precise searches, inspect caller/callee relationships, and conduct impact analysis before making code changes, moving beyond simple file/text retrieval.

Problems Solved

1. Pain Point: It eliminates the "knowledge cold-start" and "reinventing the wheel" problem in multi-agent systems. Traditional agent interactions are ephemeral and isolated, forcing repeated explanations of project context, re-reading of documentation, and rediscovery of working solutions, leading to wasted tokens, unstable outputs, and slower iteration.
2. Target Audience: The primary users are development teams and organizations building and scaling AI agent systems. This includes AI Engineers integrating memory into agent frameworks, DevOps/SRE Teams managing agent deployments, Product Teams using agents for complex workflows, and Technical Leads aiming to standardize and accumulate institutional knowledge across their agent fleet.
3. Use Cases: Essential scenarios include: onboarding a new agent into an existing project with historical context; sharing a proven troubleshooting Skill from a senior agent to a junior one; enabling a coding agent to understand codebase dependencies before refactoring; and maintaining a living wiki of product decisions that all research and analysis agents can access.

Unique Advantages

1. Differentiation: Unlike simple chat history loggers or generic RAG (Retrieval-Augmented Generation) systems, TencentDB Agent Memory provides governed, shareable, and structured memory assets. While standard RAG answers "what can be found?", this system also answers "who can use it, which version is valid, and which agent should receive it." It moves from passive storage to active team resource management.
2. Key Innovation: Its core innovation is the multi-layer memory distillation pipeline combined with a unified asset management and ACL (Access Control List) system. The layered approach (L0-L3) allows for efficient, context-aware retrieval. The unified Memory Hub with granular controls (private, team, restricted, agent-bound) enables portable, framework-agnostic memory that respects privacy and operational boundaries, a significant step beyond a single, global memory pool.

Frequently Asked Questions (FAQ)

1. How does TencentDB Agent Memory differ from simply using a vector database for chat history? While it uses vector search, it is a full memory management system. It adds structure by distilling conversations into layered assets (Chat Memory), extracts executable workflows (Skills), builds linked documentation (Wiki), and creates code dependency graphs (Code-Graph). It also includes access controls, versioning, and a team management panel, which a raw vector database lacks.
2. What AI agent frameworks is TencentDB Agent Memory compatible with? The system is designed to be framework-agnostic. It currently provides direct SDK integration and has demonstrated compatibility with OpenClaw and Hermes Agent frameworks. Its architecture, using a central Memory Hub and Proxy, is built to support broader cross-framework migration and integration.
3. Can I import my existing data, like old chat logs or code repositories, into TencentDB Agent Memory? Yes, the system supports cold-start import. You can directly import existing codebases (for Code-Graph), documents (for Wiki), and past agent conversation sessions. The system will automatically process these to create the corresponding memory assets, turning past learning costs into reusable team knowledge.
4. Is the memory shared between all agents by default, and how is privacy managed? No, memory is not globally shared by default. The system employs a granular ACL model. New Chat Memory and Skills are private to the owner. Sharing is an explicit action. Visibility levels include team (all team members), restricted (precise User/Role/Agent ACLs), and agent (for equipping specific agents). This allows experience sharing without privacy leakage.
5. What are the deployment options for TencentDB Agent Memory? It is designed for flexible deployment. The primary method is a one-command launch using Docker Compose (./start-all.sh) that spins up all core services (memory-core, memory-hub, proxy). It can also be deployed in standalone configurations, such as running only the Memory Hub service, depending on your existing infrastructure and integration needs.