Author: Tencent Cloud Database Team (腾讯云数据库团队)
Open-source project: TencentDB Agent Memory
Completing a task with an AI agent often produces more useful knowledge than the final deliverable captures.
Writing a technical article requires agreement on product terminology, feature boundaries, the intended audience, and how claims should be expressed. Fixing a bug requires understanding business constraints, the investigation process, and which approaches have already been tried. When the task is finished, that knowledge is scattered across conversations, documents, and code. Switch clients, start a new session, or hand the work to another agent, and someone has to explain it all again.
Agents are getting faster at executing tasks. Transferring experience still takes human time.
TencentDB Agent Memory has added support for the DeepSeek Harness (DSH) desktop client to help connect that handoff. Developers can associate existing memory assets with a DSH session within their authorized access scope, reusing project context, business constraints, and task methods accumulated by other connected agents.
The integration gives a team's existing experience another place to be used. It also raises a practical question: when tools keep changing, how can that experience stay with the team?
From a conversation to assets the next task can use
A long conversation can contain confirmed facts, tentative guesses, rejected approaches, and a method that eventually worked. Before another agent can take over, that material needs to be organized.
TencentDB Agent Memory uses Memory Hub to manage four types of assets:
- Chat Memory preserves information and experience from conversations, helping subsequent tasks retain project context, business constraints, and earlier decisions.
- Skill captures reusable task methods, including checks and workflows that have already worked.
- LLM-Wiki organizes document knowledge, giving tasks access to product information, designs, and specifications.
- CodeGraph represents code relationships, helping agents understand code structure, dependencies, and engineering context.
Assets can be associated with designated agents. Memory Hub manages their ownership, versions, and visibility. Developers can also import existing documents, code repositories, and historical sessions to give a new agent an initial body of knowledge and experience.
Decoupling assets from a specific agent framework makes it possible to reuse the same accumulated knowledge across connected tools. Clients provide the working interface, while the team continues to maintain its own assets.
A technical writing example: handing agreed guidance to the next agent
The mechanism applies to development work and suggests a way to organize collaboration on technical content.
Consider an article explaining how to integrate a new release. Research gathers official documentation. Drafting establishes what the product supports. Review corrects wording that could mislead readers. That work remains useful when the team later creates a tutorial, translates the article, or writes about another update.
A team could organize assets around that workflow as follows:
- A research agent gets access to product documentation and confirmed information, keeping supporting sources available as it gathers material.
- A writing agent gets terminology guidelines, audience context, and the article production workflow, reducing the need to establish the same guidance again.
- A review agent gets feature boundaries, previous review comments, and a checklist, helping it check whether claims go beyond confirmed capabilities.
This is an illustrative application scenario. Its practical results need to be evaluated in real tasks. Writing guidelines can be organized as document knowledge, and a review process that works can be captured as a Skill. If an article discusses a code implementation, the relevant CodeGraph can be associated as well.
Different roles need different assets. Research material, drafting constraints, and review records should be assigned according to the task so irrelevant information does not crowd the context.
With the DSH desktop integration, users can associate those existing assets from a new working interface and build on earlier work in the next task.
Connecting the DSH desktop client to team memory
This integration uses the Memory proxy service. Before entering settings in the desktop client, prepare:
- A Memory proxy service with the DSH model upstream configured and session initialization enabled.
- A
user_keyfor application use. - The complete DSH integration endpoint supplied by your deployment.
Once those prerequisites are ready, configure the desktop client in three steps.
Step 1: Configure the model connection. Open DSH's Settings → Models and edit the DeepSeek (deepseek-official) card. Enter the application user_key supplied by your platform. Expand Custom Settings within the card, enter the complete integration endpoint, and save. The labels here are English translations of the controls described in the Chinese setup guide; their wording may differ in your client.
Step 2: Start a new conversation. Send a message using the configured model to enter the session initialization flow.
Step 3: Associate team assets as needed. Choose whether to associate team assets for the current task, then follow the initialization prompts. Asset organization, visibility, and agent associations can be managed in Memory Hub.
After configuring the client, use a narrowly scoped task to check the connection. For example, check whether a new session can use the associated project context, then whether it follows an established business constraint. This makes it easier to determine whether the assets are actually available to the current work.
For deployment and full configuration details, see the installation guide or its Chinese version.
Reusing experience requires clear boundaries
Reuse across tools depends on two conditions: the assets are ready, and the current agent is authorized to use them.
Teams therefore need to maintain asset ownership, versions, and visibility. Personal drafts, client information, and shared team guidelines may require different access boundaries. Updated product rules also need to be distinguished from older conclusions.
Team leads need to decide which experience can be shared, which assets belong with each role, and how outdated material should be updated. Users need to check that the proxy service, model upstream, and asset-sharing scope fit the work they are doing.
Memory can supply historical evidence and reusable methods. Final deliverables still need review. Release descriptions and integration tutorials, in particular, should reflect the current documentation, and their instructions should be checked in the target environment.
Keeping work moving when tools change
Alongside DSH, TencentDB Agent Memory provides integration support for clients including Claude Code, CodeBuddy, Codex, WorkBuddy, OpenCode, Hermes, and OpenClaw. Each client has its own configuration documentation.
As work spreads across IDEs, command-line tools, and desktop applications, teams will switch agents more often. Preserving experience as assets that can be managed and associated with agents may reduce the cost of repeatedly explaining context and rediscovering methods at each transition.
The DSH desktop integration adds another entry point to that approach. The next questions are practical: can the same experience be used correctly in a new task, and can it be expanded and corrected after use?
Let each task leave behind experience that the next task can use.
Open-source project: TencentDB Agent Memory
Release history: CHANGELOG
Top comments (0)