Skip to content
TencentDB Agent Memory: Team Memory Infrastructure for AI Agents

TencentDB Agent Memory: Team Memory Infrastructure for AI Agents

TencentDB Agent Memory provides a team memory infrastructure for AI agents, organizing conversations, documents, and code into four asset types—Memory, Skill, Wiki, and CodeGraph—with unified metadata, context assembly via JEV model, and experience extraction from execution records, reducing token usage by 39.7% and task cost by 13.8% in database operations evaluations.

Smart Era Software Development
Smart Era Software Development
Smart Era Software Development

From Delivery Problems to Solution

As AI agents penetrate R&D and office workflows, teams face knowledge reuse challenges: project context must be repeatedly explained, design constraints get lost in handovers, and similar problems recur across tasks. Teams need a cross-task knowledge and experience system to give agents consistent execution grounding.

In an internal project iteration, development took 2 weeks while system testing took 8 weeks. This shows that after point-execution efficiency improves, teams still need better constraint handoff, system boundary identification, and historical experience reuse to translate execution gains into overall delivery improvement.

Figure 2: From personal memory to team memory, providing organizational context for agent execution
Figure 2: From personal memory to team memory, providing organizational context for agent execution

TencentDB Agent Memory addresses three execution gaps:

Constraint handoff: Same-source assets assembled by role. Design gets requirement background, development gets interface contracts, testing gets exception conventions. The product supplies role-specific context while keeping key facts, constraints, and versions consistent, reducing repeated explanations and handoff omissions.

System boundaries: Combining document semantics with code relationships. Wiki helps agents understand module responsibilities and interface contracts; CodeGraph provides call, dependency, and impact analysis, supporting task verification across relevant system boundaries.

Experience reuse: On-demand invocation of historical conclusions and mature methods. Memory stores historical conclusions with applicability conditions; Skill encapsulates handling methods into SOPs and checklists for later tasks to invoke.

Figure 3: Three execution gaps and corresponding Team Memory mechanisms
Figure 3: Three execution gaps and corresponding Team Memory mechanisms

These capabilities form a full-lifecycle experience reuse mechanism: before execution, acquire team knowledge and constraints; during execution, produce new results and evidence; after execution, precipitate them as assets for subsequent tasks. Agents can thus continuously reuse accumulated team knowledge and methods.

Asset Foundation and Responsibility Boundaries

Team knowledge resides in conversations, documents, code, and workflows. TencentDB Agent Memory organizes it into four asset types to serve different needs—context understanding, decision reference, and task execution.

Memory: Retains background, decisions, and conclusions. Facts, constraints, and experiences from sessions are layered to help agents restore work context and reuse existing judgments.

Skill: Reuses methods, SOPs, and checklists. Verified workflows are organized as skill assets, allowing tasks to directly call mature methods and reduce repeated exploration.

Wiki: Organizes documents and knowledge semantics. Existing documents are structured into knowledge pages with associations, giving agents retrievable, understandable project knowledge.

CodeGraph: Identifies calls, dependencies, and impact scope. A relationship index built from code structure helps agents determine change scope and supports development and verification.

Figure 4: Four asset types share a unified metadata layer managing identity, permissions, versions, and evidence
Figure 4: Four asset types share a unified metadata layer managing identity, permissions, versions, and evidence

The four asset types share a unified metadata layer recording owning team, applicable scope, source evidence, valid version, and usage. Members and agents share assets within permission scope; teams can trace asset origins and their usage in specific tasks.

The product architecture defines clear component responsibilities: Memory Core manages assets, permissions, and versions; Memory Proxy assembles context for the current task and records model usage and execution traces; Memory Panel provides configuration, viewing, and approval entry points. Together they support asset management and task execution.

Pre-execution: On-demand Context Assembly

As knowledge assets accumulate, selecting relevant content per task becomes key to context management. TencentDB Agent Memory forms a demand list based on team, repository, task, and role, then recalls corresponding assets.

When fixing issues, the product prioritizes historical constraints, call relationships, and fix checklists; when writing designs, it prioritizes requirement background, rejection reasons, and review methods. The same asset pool generates task-appropriate context, keeping execution grounding aligned with current goals.

Figure 5: From scenario identification to injection list, making asset selection vary with task
Figure 5: From scenario identification to injection list, making asset selection vary with task

For budget management, the product first retains task constraints and key decisions, then code structure, impact scope, and execution methods; general background knowledge is queried via tools on demand. Assets are selected as complete entries, preserving the applicability conditions that conclusions depend on, allocating context budget to information directly affecting task outcomes.

A JEV judgment model further filters assets: the request side judges whether candidate Skills and Memories need injection; the generation side judges whether memory needs updating. On-demand selection reduces overhead from irrelevant context and duplicate generation.

Each injection records asset ID, version, hit reason, budget consumption, and usage logs for later optimization. Optional memory timeout can degrade to non-injection; permission checks and mandatory constraints remain effective, balancing task responsiveness and access control.

Post-execution: Reliable Experience Feedback

After task completion, TencentDB Agent Memory extracts experience from real execution records, turning task outcomes into reusable team assets.

Experience extraction focuses on final decisions and rejection reasons, boundary constraints, localization conclusions, reproduction paths, and verified methods and checklists. Notably, design rejection reasons record the conditions behind decisions, providing reference for future designs.

Figure 6: Extracting experience from execution records, passing through candidate, review, and archiving into team sharing
Figure 6: Extracting experience from execution records, passing through candidate, review, and archiving into team sharing

The product synchronously records facts in the main request and completes experience extraction asynchronously. Extraction results form structured candidate assets with session and event sources; after approval they enter shared asset management. Duplicate confirmations, temporary outputs, and process noise are filtered out at this stage; personal preferences stay in personal memory.

Shared asset permissions, versions, validity periods, and usage feedback are unified under lifecycle management. Event deduplication and version verification support task recovery; archiving and rollback support asset maintenance, ensuring knowledge stays updated and usage records traceable.

The product supports cross-agent-client asset sharing. Via the unified asset metadata layer and Memory Hub , when members switch clients, models, or invocation methods, existing asset identity, permissions, versions, and source information are preserved, reducing duplicate construction.

Figure 7: Team assets support reuse across clients and models
Figure 7: Team assets support reuse across clients and models

Evidence, Retrospective, and Next Steps

Product value appears in concrete workflows: reducing repeated background explanations, enabling tasks to call mature methods, controlling context and memory generation costs, and achieving continuous experience reuse across multiple members and agents.

Evaluation data demonstrates cost benefits of on-demand asset use. In database operations evaluations , primary model missed input tokens dropped from 24.96M to 15.04M, a 39.7% reduction . In memory generation experiments, the JEV on-demand generation strategy reduced total task cost by 13.8% compared to always generating, including JEV judgment cost. Analysis of the top 50 longest sessions showed ~ 19% cost reduction .

The team calibrated product design using 10,000+ internal real sessions , combined with code versions, test cases, and execution traces, evaluating around task quality, cost, and delivery metrics. This feedback continuously feeds asset selection and governance strategies, driving product iteration.

Based on production practice, the team established design principles: select assets by demand and budget; complete knowledge extraction asynchronously; review candidate assets before sharing; retain source tracing and version rollback capabilities. These mechanisms support continuous accumulation and reuse of experience assets under explicit permission and version management.

Figure 8: Task execution, governance precipitation, process reuse, and metric feedback form a continuous team experience accumulation loop
Figure 8: Task execution, governance precipitation, process reuse, and metric feedback form a continuous team experience accumulation loop

Next phase: further optimize experience extraction, cross-project reuse, and recall evaluation; combine test left-shift, design, and subsequent execution collaboration to deepen knowledge asset participation in delivery processes.

Members directly access team-accumulated methods and judgments; agents gain background, constraints, and checklists for task execution. Facts and evidence produced by tasks continuously feed back, providing basis for organizational experience accumulation and process optimization.

TencentDB Agent Memory is now open source. Visit the GitHub repository homepage for deployment guides and agent integration methods to understand and deploy the team memory service.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

knowledge managementAI agent infrastructurecontext assemblyTencentDB Agent Memoryteam memoryJEV modelAI-native infrastructureexperience extraction
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.