TencentDB Agent Memory v2.0: A Team-Level Memory Hub for AI Coding Agents
In this article
Tencent Cloud released TencentDB Agent Memory v2.0 on August 3, 2026, publishing it under an MIT license with a one-command Docker deploy across linux/amd64 and linux/arm64. The release addresses a structural gap that anyone running multi-agent coding workflows has already felt: every new session re-explains context that a previous session already worked out. The system's answer is to convert conversations, documents, and code into four typed, versioned, permissioned memory assets that any authorized agent on the team can read — without requiring the human to repeat themselves or the system to leak anything marked private.
The timing matters because the security surface of AI coding agents is widening, and ungoverned, flat memory stores are a significant part of that exposure. TencentDB Agent Memory v2.0 ships governance as a first-class primitive, not an afterthought.
Four Assets, One Asset Model
The system defines four memory asset types, each registered under a uniform ownership, version, status, and visibility schema:
- Chat Memory captures preferences, decisions, facts, and interaction history from conversations.
- Skill distills completed tasks into reusable procedures, carrying versions, resource files, trigger boundaries, execution steps, and validation rules.
- Wiki converts documents into structured pages with a link graph, drawing on Andrej Karpathy's concept of an LLM-maintained knowledge base.
- CodeGraph indexes symbols, files, call relationships, and impact paths across a repository.
Because all four register identically as Memory Assets, sharing rules and retrieval behavior apply uniformly — a Skill and a Wiki page behave the same way from the governance layer's perspective.
| Asset Type | Source Material | Key Contents | Primary Use Case |
|---|---|---|---|
| Chat Memory | Conversation turns | Preferences, facts, decisions, interaction history | Session continuity, persona persistence |
| Skill | Completed tasks | Versions, resource files, trigger boundaries, execution steps, validation rules | Reusable agent procedures, runbooks |
| Wiki | Documents | Structured pages, link graph | Product docs, onboarding pages, code review standards |
| CodeGraph | Repository | Symbols, files, call relationships, impact paths | Refactor impact analysis, codebase onboarding |
Distillation Pipeline and Retrieval Budget
Chat Memory is not stored flat. Conversations are saved as L0 raw turns, then an asynchronous pipeline refines them into L1 Atom, L2 Scenario, and L3 Core/Persona layers. On retrieval, L2 and L3 provide a fast context bootstrap. When specific facts are required, the system falls back through BM25 plus vector retrieval plus Reciprocal Rank Fusion (RRF) down to L1 and L0. Results are capped by item count, character budget, and timeout to prevent memory from crowding out the context window — a practical constraint that distinguishes this from naive RAG injection. For teams thinking about what the agentic era demands of data infrastructure, this kind of budgeted, layered retrieval is exactly the architecture that scales beyond single-session demos.
Governance as the Real Differentiator
Standard RAG resolves what can be retrieved. The Memory Hub also resolves who may use it, which version is valid, and which agent receives it. Visibility tiers are private (owner-only, explicitly not readable by team admins), team, and restricted, with an agent target for direct equipping. New Chat Memory and Skills default to private, making sharing a deliberate act rather than the default. Binding narrows by team, user, agent, and visibility before retrieval runs — ACL-first, search-second.
Memory Proxy exposes both Anthropic and OpenAI protocols via /claude-code/<spaceId>/v1/messages and /v1/chat/completions. On the first turn, sessionInit uses Claude Code's native AskUserQuestion tool to select team, agent, and task. Every subsequent turn injects the agent's L2/L3 memory, matched skills, and knowledge into the system prompt before forwarding upstream. Default ports per INSTALL.md are 8420 (Memory Core), 8125 (panel), 8424 (knowledge service), and 8096 (proxy). Supported integrations include OpenClaw, Hermes, Claude Code, CodeBuddy, and direct SDK use via official TypeScript and Python packages. The 2.0.0 release adds Skill forced archiving, scheduled CodeGraph repository sync, system-admin asset management, English/Chinese panel switching, and a Cost Guard feature that routes specific agents to cheaper models.
On benchmarks, the project self-reports PersonaMem accuracy improving from 48% to 76% — a 59% relative gain — with no independent reproduction published yet. Private-repository CodeGraph and automated memory routing are also noted as still being refined, which teams in regulated environments should factor into any deployment decision.
The broader signal here is that persistent, governed memory is transitioning from a research concern into shipping infrastructure. As multi-agent coding stacks grow more complex — and as the security implications of those stacks become better understood — the teams that will move fastest are the ones whose agents share learned context without requiring human relay. An MIT-licensed, self-hosted hub deployable with a single Docker command is a meaningful step toward making that the default rather than the exception.