TencentDB Agent Memory

Overview

TencentDB Agent Memory takes a different approach from individual-agent memory systems. Instead of giving one agent its own memory store, it provides a team-level hub where multiple agents share governed, structured knowledge. Conversations become Chat Memory. Documentation becomes LLM-Wiki. Code becomes CodeGraph. Agent workflows become reusable Skills. Each asset type is stored, versioned, and access-controlled independently, and agents are equipped with specific memory loadouts based on their role.

Key Capabilities

Four memory asset types: Chat Memory retains user preferences, facts, decisions, and interaction history, organized in layers from L0 (raw conversations) through L1 (extracted atoms), L2 (scenarios), to L3 (personas). Skills capture reusable workflows extracted from agent tool calls, complete with execution steps, resource files, and validation rules. LLM-Wiki converts product documentation into interconnected, auto-indexed pages with document-to-document link graphs. CodeGraph indexes code symbols, files, call relationships, and impact paths for dependency-aware code reasoning.

Team-level governance: Unlike personal memory systems, TencentDB Agent Memory provides role-based access control across all assets. Visibility tiers (private, team, restricted, agent) control who can read each memory asset. Skills start as private to the creating agent and can be promoted to team-shared after review. System admins manage global policy while team-level roles handle day-to-day access.

Agent equipping: Memory assets are bound to agents explicitly, not dumped wholesale. A research agent gets market analysis Wiki and data-gathering Skills. A build agent gets product Wiki and CodeGraph. A reviewer gets incident history and release checklists. Priority and retrieval mode are configurable per agent, so different roles pull different context from the same memory hub.

Cold-start capability: Teams do not start from zero. Existing codebases are imported and automatically indexed into CodeGraph. Documentation repositories are structured into LLM-Wiki. Prior agent conversation histories are ingested and processed into Chat Memory and Skills. The continuous learning pipeline then refines these assets asynchronously as new interactions occur.

Use Cases

Development teams deploy TencentDB Agent Memory when running multiple specialized agents that need shared context — a scout agent researching, a builder agent implementing, and a reviewer agent validating, all drawing from the same governed knowledge base. Enterprise teams use the access control model to partition sensitive assets while allowing cross-team knowledge sharing where appropriate. Teams migrating between agent frameworks use the proxy adapter to maintain memory continuity across tool changes.

Considerations

The three-service architecture (Memory Core, Memory Hub, Proxy) requires Docker for deployment. The web management panel runs at localhost:8125. PersonaMem benchmark shows 76% accuracy with memory enabled versus 48% baseline — a 59% improvement in correctly applying user context across extended interactions. The project is MIT-licensed and growing rapidly (22K+ stars in under four months). The framework-agnostic proxy adapter means switching agent platforms does not require re-building the memory store.

Who It’s For

TencentDB Agent Memory is for teams running multi-agent workflows that need governed, shared memory across different agent roles and frameworks. It is particularly valuable for organizations where multiple agents collaborate on complex tasks and need structured knowledge sharing with access control — not just individual agent recall, but team-level knowledge management.