Back to feed
News Story
APriority74
InfoQ
1 sources

Tencent Cloud Open-Sources TencentDB Agent Memory to Give AI Agents Persistent Memory

Tencent Cloud has released TencentDB Agent Memory 2.0.0 beta, marking the first public open-source of its agent memory solution. The solution uses four types of memory assets and an asset catalog to help AI agents reuse team experience across sessions, improving task efficiency. In a public benchmark, accuracy improved from 48% to 76%, demonstrating the significant impact of memory structure on agent performance.

SynthePulse Insight · AI deep readingMembers

TencentDB Agent Memory 2.0.0 beta Open Source: Let Agents Continue from Team Memory

Version 1 · 1 source

Tencent Cloud open-sources TencentDB Agent Memory, aiming to solve the problem that AI Agent memory cannot be reused in production environments, enabling cross-session and cross-agent memory inheritance through four types of assets and an asset catalog.

  • TencentDB Agent Memory released 2.0.0 beta in July 2026, its first public open-source release, including Memory Core, Memory Hub, Memory Proxy, and SDKs.
  • The framework abstracts team memory into four types of atomic assets: Chat Memory, LLM-Wiki, Code Graph, and Skill.
  • The asset catalog describes identity, scope, trust, lifecycle, governance, usage, and other information; the Proxy performs recall and permission filtering based on task relevance.
Open section navigationProblem: Every Agent Conversation Starts from Scratch

Problem: Every Agent Conversation Starts from Scratch

Over the past year, AI Agents have been rapidly entering real development scenarios, capable of reading code, using tools, running tests, and advancing complete requirements. But the closer to production, the sharper the problem: every conversation feels like starting over, personal experience doesn't accumulate, and team context, technical trade-offs, and task processes cannot be reused. What blocks Agents from going from 'usable' to 'good' is not single-shot reasoning, but memory.

Development teams never lack materials, but information is scattered across wikis, code repositories, issues, and chats, making it difficult for Agents to accurately recall at the start of the next task. In a bug fix, an Agent might have determined that a certain compatibility branch cannot be deleted and discovered the call chain of related modules, but if these conclusions only stay in the current session, the next time it encounters a similar problem, it will rescan, re-reason, and even step into the same pitfalls again.

Free for now

Read the full analysis

2 more sections of analysis, plus the full takeaway

Loading

Credibility boundary

The information in this article mainly comes from official Tencent Cloud releases, republished by InfoQ. The evaluation data is officially claimed and has not been independently verified.

Sources for this version

  1. 别让 Agent 每次都重新开始InfoQ

Primary report

InfoQ

Primary source