TencentDB Agent Memory: How Tencent’s Layered Memory System Makes Long‑Running AI Agents Cheaper and More Reliable

The Four‑Layer Long‑Term Memory Architecture

Tencent’s design organizes long‑term memory into four progressive layers that transform raw interactions into structured knowledge.

L0: Raw Dialogue Layer
Stores complete conversation records and task interactions exactly as they occurred.

L1: Atomic Memory Layer
Extracts structured facts from those interactions—such as user preferences, constraints, or conclusions from previous steps.

L2: Scenario Summary Layer
Aggregates memories related to a particular task or scenario, enabling the agent to recall patterns across similar workflows.

L3: User Profile Layer
Distills long‑term behavioral patterns and preferences into a compact user profile.

The effect is a gradual transformation from raw conversations into reusable structured knowledge. Over time, agents can reuse previous experiences rather than recomputing them from scratch.

The Core Innovation: Context Offloading + Mermaid Task Canvas

The system’s biggest efficiency gain comes from how it handles short‑term working memory during long tasks.

Context Offloading

After an agent performs a tool call—such as fetching a webpage or executing code—the full output is stored outside the prompt in external storage. Only a high‑density summary or reference remains in the model context.

This prevents large tool outputs, logs, or documents from permanently occupying prompt space.

Mermaid Task Canvas

Instead of storing long textual histories, Tencent represents task progress using a structured task graph written in Mermaid, a text‑based diagram language widely used in developer documentation.

The canvas acts like a navigation map for the agent:

• nodes represent task steps
• edges represent dependencies
• each node contains a short summary or state marker

Because the model only needs to reason about the task structure rather than every raw message, it can track complex workflows with far fewer tokens.

Tencent describes the difference with a simple analogy: logs record everything, but maps help you navigate. The Mermaid task canvas functions as that map for the agent.

Adaptive Compression Based on Context “Water Level”

TencentDB Agent Memory also compresses context dynamically as the prompt fills up. The system monitors how much of the context window is being used and applies different compression levels.

Typical thresholds include:

• Real‑time summaries (L1 compression): tool outputs are summarized immediately after execution.
• Task‑canvas updates (L2 compression): the Mermaid task map is updated asynchronously to capture workflow structure.
• Deep compression (L3): when context usage rises to around 80% or higher, older messages are aggressively compressed or removed.

If usage approaches critical levels (around 95%), the system triggers emergency compression to reduce the context load again.

Reported Benchmark Results

Tencent reported several performance improvements when integrating Agent Memory into agent frameworks. These results come from internal experiments and should be interpreted as vendor‑reported results rather than independent benchmarks.

Key reported results include:

WideSearch benchmark

• Pass rate: 33% → 50% (51.52% relative improvement)
• Token usage: 221.31M → 85.64M (61.38% reduction)

SWE‑bench

• Success rate: 58.4% → 64.2%
• Token usage: 3474.1M → 2375.4M (33.09% reduction)

AA‑LCR benchmark

• Success rate: 44.0% → 47.5%
• Token usage: 112.0M → 77.3M (30.98% reduction)

PersonaMem benchmark

• Accuracy: roughly 48% → 76% after adding the memory system.

Tencent also reported tests across 1,540 tasks spanning code generation, web search, document analysis, and long multi‑step workflows, with overall task completion improving 12–35% while token consumption dropped 33–64%.

What Changed Between the April Launch and the May 14 Release

TencentDB Agent Memory was introduced earlier in 2026, but the focus evolved between releases.

April launch

• Introduced the long‑term memory system with the four‑layer architecture
• Demonstrated improved performance on PersonaMem benchmarks
• Focused primarily on persistent memory across sessions

May 14 open‑source release

• Released the full stack as open source under the MIT license
• Added the short‑term memory compression system for long tasks
• Highlighted the Context Offloading + Mermaid Task Canvas mechanism

In short, the earlier launch emphasized persistent memory, while the open‑source release focused on solving context‑window overload during active agent tasks.

Framework Integrations

Tencent says the system already integrates with several agent frameworks.

Examples include:

• OpenClaw, where it can function as a memory enhancement plugin
• Hermes Gateway / Hermes Agent, with Docker deployment support for Hermes Gateway 0.3.4 or later

These integrations allow developers to add memory compression and long‑term memory to existing agent architectures without redesigning the entire system.

Why This Matters for the AI Agent Race

As AI agents move from demos to real applications—coding assistants, research agents, and enterprise workflow automation—the economics of context windows become a major bottleneck. Long chains of tool calls can rapidly inflate token usage and degrade reasoning quality.

Tencent’s approach tackles two problems simultaneously:

• Cost: reducing token usage significantly lowers operational costs for long tasks.
• Reliability: structured task memory helps agents maintain direction across complex workflows.

If these improvements hold up in broader testing, systems like TencentDB Agent Memory could become an important infrastructure layer for autonomous AI agents.

For now, though, the benchmark improvements remain vendor‑reported results, and wider independent validation will determine how well the approach performs across different models and agent frameworks.