⚛️量子位•Stalecollected in 56m
Tencent open-sources Agent memory tech, cuts tokens by 61%
💡A practical open-source tool to drastically reduce LLM inference costs for agents.
⚡ 30-Second TL;DR
What Changed
Open-source memory optimization framework for AI Agents
Why It Matters
This release provides developers with a cost-effective way to manage long-term memory in LLM-based agents. It lowers the barrier for building complex, stateful AI applications.
What To Do Next
Integrate Tencent's memory management framework into your existing agentic workflows to reduce API costs.
Who should care:Developers & AI Engineers
Key Points
- •Open-source memory optimization framework for AI Agents
- •Reduces token usage by up to 61%
- •Improves task execution success rate by 51%
🧠 Deep Insight
Web-grounded analysis with 17 cited sources.
🔑 Enhanced Key Takeaways
- •The open-source solution, named TencentDB Agent Memory, provides both short-term memory compression and long-term personalized memory capabilities for AI Agents.
- •TencentDB Agent Memory implements a "four-layer progressive memory system" designed to transform fragmented conversations into structured facts, contextual cognition, and personalized profiles.
- •The short-term memory component utilizes "Context Offloading + Mermaid Task Canvas" technology, which offloads complete information to external storage while retaining key states and execution paths in a structured task graph.
- •The long-term memory feature distills fragmented conversations into structured personas and scenes, moving beyond simple flat vector storage and preserving a clear path from high-level abstraction back to ground-truth evidence.
- •When integrated with OpenClaw, TencentDB Agent Memory has demonstrated a 59% relative improvement in PersonaMem accuracy and a 51.52% relative increase in task pass rates on benchmarks like WideSearch.
📊 Competitor Analysis▸ Show
| Feature/Framework | TencentDB Agent Memory | Mem0 | Zep | Letta (MemGPT) | Hindsight |
|---|---|---|---|---|---|
| Memory Class | Symbolic short-term, layered long-term (personas, scenes) | Personalization + some institutional (vector + graph) | Both, strongest on temporal institutional knowledge (temporal knowledge graph) | Both (tiered, OS-inspired, self-editing) | Both (personalization + institutional) |
| Architecture | Context Offloading + Mermaid Task Canvas for short-term; four-layer progressive system for long-term | Dual-store (vector DB + knowledge graph) | Temporal knowledge graph (Graphiti) | Three-tiered (Core, Recall, Archival memory) | Multi-strategy hybrid |
| Open Source | Yes | Yes (Apache 2.0) | Graphiti (open-source engine) | Yes (Apache 2.0) | Yes (MIT) |
| Key Strengths | Significant token reduction (up to 61%), improved task success (up to 51%), structured memory for personas/scenes | Widely adopted, framework-agnostic, strong personalization, well-funded | Excellent for temporal reasoning and long-running sessions, tracks entity changes over time | Mimics OS memory management, agents control their own memory, effectively unlimited memory | High accuracy on LongMemEval (91.4%), designed for extracting lessons and building domain understanding |
| Benchmarks (Accuracy/Tokens) | WideSearch: +51.52% pass rate, -61.38% tokens; PersonaMem: +59% accuracy | LOCOMO: 67.13% accuracy, ~1,764 tokens/conversation (vs. 26,031 for full-context) | LongMemEval: 63.8% (claimed 75.14% on LOCOMO) | LOCOMO: Weak (5-9 points) | LongMemEval: 91.4% overall accuracy |
🛠️ Technical Deep Dive
- System Name: TencentDB Agent Memory
- Core Architecture: A "four-layer progressive memory system" that processes information from original conversations to user profiles.
- Short-Term Memory Mechanism (Symbolic Memory): Employs "Context Offloading + Mermaid Task Canvas" technology. This involves offloading complete, heavy tool logs and other information to external storage. Key states and execution paths are condensed and retained in a structured task graph using Mermaid symbols, allowing the Agent to maintain a lightweight context.
- Long-Term Memory Mechanism (Layered Memory): Distills fragmented conversations into structured personas and scenes, moving beyond flat vector storage. It preserves a clear, auditable path from high-level abstractions (judgments, directions) back to ground-truth evidence and precision in lower layers.
- Integration: Seamlessly integrated into Tencent Cloud products like Tencent Cloud Lighthouse and ClawPro through plugins, and tested with OpenClaw.
- Performance Measurement: Results are measured over continuous long-horizon sessions, such as SWE-bench, which involves 50 consecutive tasks per session to simulate real-world context accumulation pressure.
🔮 Future ImplicationsAI analysis grounded in cited sources
Widespread adoption of TencentDB Agent Memory could accelerate the development of more capable and cost-efficient AI agents.
By significantly reducing token consumption and improving task success rates, it lowers operational costs and enhances agent reliability, making advanced AI agents more practical for diverse applications.
The "four-layer progressive memory system" could become a standard architectural pattern for future AI agent memory solutions.
Its ability to transform fragmented conversations into structured facts and personalized profiles addresses a core challenge in AI agent long-term memory, offering a robust model for context management.
Tencent's open-sourcing strategy will intensify competition in the AI agent memory framework market.
Providing a high-performance, open-source solution from a major tech company will push other developers and companies to innovate further and potentially open-source their own advanced memory solutions to remain competitive.
⏳ Timeline
2025-07-02
Tencent released Hunyuan-A13B, an open-source AI model with a 256K token context window, demonstrating early commitment to large context and open-source AI.
2026-03-02
Tencent Cloud CodeBuddy's Memory feature was updated, enabling AI agents to automatically remember user preferences and important information for personalized assistance.
2026-04-03
Tencent Cloud officially launched the 'TencentDB Agent Memory' long-term memory service, addressing the lack of persistent memory in AI frameworks.
2026-04-23
Tencent launched 'Hy3 preview,' a new open-source AI model, which included significantly improved agentic capabilities and strong performance on in-house benchmarks for OpenClaw compatibility.
2026-05-13
Tencent Cloud officially open-sourced TencentDB Agent Memory, providing symbolic short-term memory compression and layered long-term memory capabilities.
📎 Sources (17)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
- Google Search Source
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 量子位 ↗