Agent 记忆真是太特么卷了啊！ 不得不说，这个赛道越多人加入越爽啊！ Tencent AI团队花了整整6个月，就死磕一个问题：AI agent长会话里疯狂丢上下文。 他们最后把一套记忆系统做完，直接开源了。 我看完他们的分享，最大的感受是，99%的人还在卷上下文长度，真正把agent拉回正轨的，是这三招硬核操作。 第一招，实时压缩过期上下文。 直接把token消耗砍掉61%。 以前动不动就爆上下文，现在中途就瘦身，agent还能保持清醒。 第二招，给agent画一张结构化的任务地图，用mermaid语法直接生成。 30多步的复杂流程里，丢轨概率大幅下降。 agent不再像无头苍蝇，它知道自己现在在哪一步，该往哪走。 第三招，专门给agent建了Persona记忆。 人格一致性从48%直接跳到76%。 它不再一会儿专业一会儿随意，回答风格和角色设定稳得一批。 这套东西不是理论，是他们真实踩坑6个月踩出来的。 Repo已经挂出来了，谁在做agent的赶紧去试。 以前总觉得agent记忆难，是因为我们把问题想简单了。 真正难的不再是存更多信息，是需要解决让它在正确的时间用正确的方式想起正确的东西。 你还在靠堆token解决agent记忆问题吗？ 项目地址：https://github.com/Tencent/TencentDB-Agent-Memory…

Tencent/TencentDB-Agent-Memory

Source: https://github.com/Tencent/TencentDB-Agent-Memory

Agents remember,Humans innovate.

Highlights · Overview · Core Technology · Features · Quick Start

English · 简体中文

✨ Highlights

TencentDB Agent Memory = symbolic short-term memory + layered long-term memory.

• Symbolic short-term memory offloads heavy tool logs and condenses them into compact Mermaid symbols, cutting token usage and improving task success.
• Layered long-term memory distills fragmented conversations into structured personas and scenes, instead of flat vector piles.

When integrated with OpenClaw, it cuts token usage by up to 61.38%, improves pass rate by 51.52% (relative), and raises PersonaMem accuracy from 48% to 76%.

These results are measured over continuous long-horizon sessions, not isolated turns. For example, SWE-bench runs 50 consecutive tasks per session to simulate the context-accumulation pressure of real-world long-horizon agents.

Overview

Memory is not about hoarding everything in the AI — it is about sparing humans from having to repeat themselves.

In practice, we constantly re-explain the same SOPs, project background, tool conventions, and output formats to the Agent. Such information should not require repetition, nor should it be indiscriminately dumped into the context.

TencentDB Agent Memory helps the Agent learn your workflows, retain task context, and reuse past experience. We reject both brute-force history accumulation and irreversible lossy summarization. Instead, we design memory as a layered system: symbolic memory for in-task information overload, and memory layering for cross-session experience.

Let the Agent remember what should be remembered, so people can focus on judgment, creation, and work that truly matters.

Core Technology: Reject Flat Storage, Embrace Layering and Symbolization

Our architecture rests on two pillars: memory layering and symbolic memory. Together they ensure Agents do not merely “remember more”, but “reason better”.

1. Memory Layering: Progressive Disclosure with Heterogeneous Storage

Traditional memory systems shred data into fragments and dump them into a flat vector store. Recall degenerates into a blind search across disconnected fragments, with no macro-level guidance.

Whether it is long-term knowledge, short-term tasks, or future skill capabilities, memory should never be flat — both its formation and its recall must be hierarchical. TencentDB Agent Memory adopts layering as its unified architectural paradigm:

• Short-term context layering. The bottom layer archives raw tool outputs (refs/*.md); the middle layer extracts step-level summaries (jsonl); the top layer condenses state into a lightweight Mermaid canvas. The Agent only needs to attend to the top-layer structure in context, and drills down to the lower layers via node_id when an error occurs.
• Long-term personalization layering. In place of flat logs, we build a semantic pyramid: L0 Conversation (raw dialogue) → L1 Atom (atomic facts) → L2 Scenario (scene blocks) → L3 Persona (user profile). The Persona layer carries day-to-day preferences; the system drills down to Atoms only when details matter.
• Skill generation layering. Layering also applies to actions. The middle layer derives common solution patterns (Scenario) from bottom-layer execution traces (Conversation), and the top layer distills reusable Skills or standard SOPs (Persona).

Heterogeneous storage and progressive disclosure. A dual-layer storage strategy underpins this architecture. The bottom layer (facts, logs, traces) is persisted in databases for robust full-text retrieval; the top layer (personas, scenes, canvases) is stored as human-readable Markdown files for high information density and white-box inspection. Lower layers preserve evidence; upper layers preserve structure.

Full traceability and lossless recovery. Compression often sacrifices traceability. TencentDB Agent Memory avoids irreversible compression by maintaining a deterministic path from high-level abstractions back to ground-truth evidence. Whether it is an offloaded error log or a distilled user preference, the system guarantees a complete drill-down path: “top-layer symbol (Persona / canvas) → mid-layer index (Scenario / jsonl) → bottom-layer raw text (L0 Conversation / refs)”.

2. Symbolic Memory: Maximum Semantics in Minimum Symbols (Mermaid Canvas)

In long tasks, the largest token consumers are verbose intermediate logs (search results, code, error traces). To address this, we combine context offloading with symbolic memory:

• Mermaid symbol graph. Instead of verbose prose or flat JSON, we encode task state transitions in high-density Mermaid syntax — precise enough for LLMs to parse, concise enough for humans to read.
• History offloading. Full tool logs are offloaded to external files; only a lightweight Mermaid task map remains in context.
• node_id tracing. The Agent reasons over the symbol graph; to verify a detail, it greps for the node_id and instantly retrieves the full raw text — cutting token cost while preserving full traceability.

graph LR
    Log["Verbose Logs<br/>(hundreds of thousands of tokens)"] -->|"1. Offload full text"| FS[("External FS<br/>(refs/*.md)")]
    Log -->|"2. Extract relations"| MMD["Mermaid Canvas<br/>(with node_id)"]
    
    MMD -->|"3. Light injection"| Agent(("Agent Context<br/>(a few hundred tokens)"))
    Agent -. "4. Recall via node_id" .-> FS
    
    style Log fill:#f1f5f9,stroke:#94a3b8,stroke-dasharray: 5 5,color:#475569
    style FS fill:#f8fafc,stroke:#cbd5e1,stroke-width:2px,color:#334155
    style MMD fill:#eff6ff,stroke:#3b82f6,stroke-width:2px,color:#1e3a8a
    style Agent fill:#fffbeb,stroke:#f59e0b,stroke-width:2px,color:#92400e

Quick Start

🎬 Demos

1. OpenClaw

1.1 Install the plugin

openclaw plugins install @tencentdb-agent-memory/memory-tencentdb
openclaw gateway restart

1.2 Zero-config to enable

Defaults to a local SQLite + sqlite-vec backend.

// ~/.openclaw/openclaw.json
{
  "memory-tencentdb": {
    "enabled": true
  }
}

Once enabled, TencentDB Agent Memory automatically handles conversation capture, memory extraction, scene aggregation, persona generation, and recall before the next turn.

1.3 Enable short-term compression (optional, requires version ≥ 0.3.4)

{
  "memory-tencentdb": {
    "config": {
      "offload": {
        "enabled": true
      }
    }
  }
}

Step 1 — Register the slot in your plugin config

Add the slots field so OpenClaw routes context-offload requests to this plugin:

{
  "plugins": {
    "slots": {
      "contextEngine": "openclaw-context-offload"
    }
  }
}

Step 2 — Apply the runtime patch

For the best results, run the patch script below. It hooks after-tool-call messages so they can be offloaded and recovered correctly:

bash scripts/openclaw-after-tool-call-messages.patch.sh

💡 The patch only needs to be applied once per OpenClaw installation. After upgrading OpenClaw, re-run the script to re-apply.

2. Hermes (Docker, requires version ≥ 0.3.4)

In addition to OpenClaw, this plugin also supports Hermes Agent. You can launch a memory-enabled Hermes with a single command:

# ============ Configuration Parameters ============
# MODEL_API_KEY    LLM API key (required) — replace with your own credential
# MODEL_BASE_URL   LLM endpoint, defaults to Tencent Cloud LKE (Large Model Knowledge Engine)
# MODEL_NAME       Model name, defaults to DeepSeek-V3.2
# MODEL_PROVIDER   Provider type: "custom" works for any OpenAI-compatible endpoint

MODEL_API_KEY="your-api-key"
MODEL_BASE_URL="https://api.lkeap.cloud.tencent.com/v1"
MODEL_NAME="deepseek-v3.2"
MODEL_PROVIDER="custom"

# ============ docker run Flags ============
# -d                          Run container in detached (background) mode
# --name hermes-memory        Container name, for later docker exec / logs / stop
# --restart unless-stopped    Auto-restart on crash or host reboot
# -p 8420:8420                Host port ↔ container port (Hermes Gateway)
# -e MODEL_*                  Inject the config parameters above as env vars
# -v hermes_data:/opt/data    Persist memory data to a named volume (survives restart)

# Build
docker build -f Dockerfile.hermes -t hermes-memory .

# Run
docker run -d \
  --name hermes-memory \
  --restart unless-stopped \
  -p 8420:8420 \
  -e MODEL_API_KEY="your-api-key" \
  -e MODEL_BASE_URL="https://api.lkeap.cloud.tencent.com/v1" \
  -e MODEL_NAME="deepseek-v3.2" \
  -e MODEL_PROVIDER="custom" \
  -v hermes_data:/opt/data \
  hermes-memory

# Verify the Gateway
curl http://localhost:8420/health

# Enter the Hermes interactive shell
docker exec -it hermes-memory hermes

The image ships with Tencent Cloud DeepSeek-V3.2 as the default. If you use this model, omit MODEL_BASE_URL / MODEL_NAME / MODEL_PROVIDER and pass only MODEL_API_KEY.

🔧 Configurable Parameters

Every field has a sensible default — it runs with zero configuration. When you want to tune, peel back the layers based on how deep you go.

🤔 Features

1. Macro Personas + Micro Facts: A Unified Drill-Down Mechanism

The biggest risk in compression is saving tokens at the cost of losing the evidence. TencentDB Agent Memory therefore does not collapse history into an irreversible summary — it preserves a clear path from high-level abstraction back to ground-truth evidence.

The upper layers carry judgment and direction; the lower layers carry evidence and precision. Short-term compression and long-term memory form a single closed loop: collapsible and expandable, abstract yet auditable.

2. White-Box Debuggability: Memory Is Not a Black Box

Most memory systems fall short here: when recall is wrong, all you see is a list of vector scores, with no way to tell where things went wrong. TencentDB Agent Memory keeps the key intermediates as readable files:

• L2 Scenario blocks are plain Markdown — open them and inspect.
• L3 Persona lives in persona.md and traces back to the Scenarios that produced it.
• Short-term task canvases are Mermaid — readable by both humans and Agents.
• Raw payloads, summaries, and nodes are linked by result_ref and node_id.

Debugging no longer means probing an opaque database — it becomes a deterministic walk along the chain “Persona → Scenario → Atom → Conversation” until the root cause surfaces.

All of these layered memory artifacts live under ~/.openclaw/memory-tdai/ — feel free to open the directory and inspect each layer for yourself.

3. Production-Ready Engineering: Not a Demo

Documentation

Community & Contributing

We welcome every kind of contribution — bug reports, feature ideas, doc fixes, benchmark reproductions, ecosystem integrations, or a Pull Request. Agent memory is far from a solved problem, and we’d love to figure it out together.

• 🐞 Found a bug or have a question? Open an issue at GitHub Issues — we respond within 24 hours.
• 💡 Have an idea to share? Start a thread in GitHub Discussions.
• 🛠️ Want to contribute code? Please read CONTRIBUTING.md first.
• 💬 Want to chat with us? Join our Discord community and talk to the early developers directly.

Roadmap

• Long-term personalized memory (L0 → L3)
• Short-term context compression (Context Offload + Mermaid canvas)
• Local SQLite backend and Tencent Cloud Vector Database (TCVDB) backend
• OpenClaw plugin and Hermes Gateway integration
• Portable memory: cross-Agent / cross-framework / cross-device import, export, and live migration
• Automatic Skill generation
• Visual debugging and memory observability dashboard

If TencentDB Agent Memory has been useful to you, please give the project a ⭐ to support us. For any suggestions, feel free to open an issue and start the discussion.

MIT © TencentDB Agent Memory Team