What is TencentDB Agent Memory
TencentDB Agent Memory is a layered memory management tool for AI Agents open-sourced by the Tencent Cloud Database team under the MIT license. Through its original L0-L3 four-layer progressive memory architecture and context offloading + Mermaid task canvas technology, the tool provides Agents with short-term memory compression and long-term personalized memory capabilities, reducing Token consumption by up to 61.38% and increasing task success rate by 51.52%. It has been adapted for mainstream frameworks such as OpenClaw and Hermes.
Main features of TencentDB Agent Memory
-
Short-term memory compression: Raw tool results are moved to external storage through context offloading, and a Mermaid task canvas is used to retain a structured task graph, so that only summaries and indexes remain in the context.
-
Long-term personalized memory: Builds a four-layer progressive architecture of L0 raw dialogue → L1 atomic memory → L2 scenario chunking → L3 user profile, enabling the accumulation of cross-session user preferences.
-
White-box traceability: All memory layers are saved as human-readable files such as Markdown, JSONL, and Mermaid, supporting layer-by-layer tracing from high-level abstractions down to the underlying raw evidence.
-
Heterogeneous storage backends: By default, it uses local SQLite + sqlite-vec for zero-dependency deployment; advanced usage supports connecting to the Tencent Cloud Vector Database TCVDB to enable BM25 + Vector hybrid retrieval.
-
Multi-framework adaptation: It has been adapted for mainstream Agent frameworks such as OpenClaw and Hermes, supporting one-click plugin installation and integration.
How to use TencentDB Agent Memory
- Install the plugin: In theOpenClawterminal, run the
openclaw plugins install @tencentdb-agent-memory/memory-tencentdbcommand to install the memory plugin. - Restart the service: After installation is complete, run
openclaw gateway restartto restart the gateway so the plugin takes effect. - Configure storage: By default, the local SQLite + sqlite-vec backend is used with no additional configuration required; for advanced use, you can configure the Tencent Cloud Vector Database TCVDB to enable hybrid retrieval.
- Verify the effect: After restarting, observe the decrease in Token consumption and task state retention in long-task Agent sessions, and trace memory layer contents through locally readable files.
Core advantages of TencentDB Agent Memory
-
Significant cost reduction and efficiency improvement: In multi-task continuous Session experiments, Token consumption was reduced by up to 61.38%, while task success rate increased by 51.52% relatively, achieving optimization in both cost and experience.
-
White-box debuggability: All memory intermediate products are human-readable files, supporting 100% layer-by-layer tracing and retrieval, avoiding information loss caused by black-box compression.
-
Zero-dependency, ready to use out of the box: By default, it uses local SQLite storage with no need for external APIs or cloud services; installation and deployment can be completed with a single command.
-
Production-grade validation: It has been validated in real production environments on four types of long-chain tasks—programming, research, document analysis, and workflow orchestration—and converges stably.
-
Progressive abstraction: The original four-layer memory architecture increased user profile accuracy from 48% to 76% in the PersonaMem evaluation, nearly 59% higher than native OpenClaw.
Project address of TencentDB Agent Memory
- GitHub repository：https://github.com/Tencent/TencentDB-Agent-Memory
Comparison of TencentDB Agent Memory with similar competing products
| Comparison dimension | TencentDB Agent Memory | Mem0 | Letta |
|---|---|---|---|
| Product positioning | Layered memory engine (short-term compression + long-term memory) | Lightweight memory layer (personalized fact storage) | Full agent runtime (self-editing memory blocks) |
| Memory architecture | L0-L3 four-layer progressive architecture | Vector storage + LLM fact extraction | Three-layer memory (core/recall/archival) |
| Short-term compression | Context offloading + Mermaid canvas | No dedicated short-term compression mechanism | Agent autonomous context management |
| White-box traceability | Full-chain human-readable files | Black-box system management | Black-box agent self-editing |
| Deployment dependencies | Local SQLite with zero dependencies | Requires an additional vector database backend | GPU support recommended |
| Integration method | OpenClaw/Hermes one-click plugin | Broad SDK integration | Requires migration to the Letta runtime |
| Open-source license | MIT | Apache 2.0 | Apache 2.0 |
| Applicable scenarios | Long-task context compression and cross-session memory | Large-scale user personalized memory | Long-term autonomous agent operation |
Application scenarios of TencentDB Agent Memory
-
Programming development assistant: Continuously remember coding standards, constraints, and progress milestones across multi-day code projects, avoiding repeated confirmation.
-
In-depth research and analysis: Maintain task state in long-chain tasks such as web search and research analysis, preventing reasoning breaks caused by the loss of intermediate results.
-
Workflow orchestration: Record execution paths and key states in multi-step automated workflows, supporting checkpoint resumption and state recovery for complex tasks.
-
Personalized customer service: Accumulate user preferences, historical requests, and profile information across sessions to provide a consistently personalized service experience.
-
Document analysis and processing: Compress historical context during paragraph-by-paragraph analysis of long documents to maintain the coherence and accuracy of the analysis logic.