Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.
Python agent-contextagent-memoryagent-pluginsagentic-ragcontext-databasedsh-pluginknowledge-basememorymemory-managementself-evolvingvector-database
docs: update benchmark results for LoCoMo and RAG (#2275)
AlexMaca55f6cb07f4
diff --git a/README.md b/README.mdindex 958c3ce3c..3702848db 100644--- a/README.md+++ b/README.md@@ -57,6 +57,66 @@ With OpenViking, developers can build an Agent's brain just like managing local***+## Evaluation Highlights++OpenViking 1.0 has been evaluated across three scenarios: long-conversation user memory, agent experience memory, and knowledge-base QA.++### 1. User Memory on LoCoMo++On the LoCoMo benchmark, OpenViking improves long-context QA accuracy while reducing both latency and token usage across multiple agent integrations:++| Integration | Accuracy | Avg. Query Time | Total Input Tokens |+| ----------- | -------- | --------------- | ------------------ |+| OpenClaw + native memory | 24.20% | 95.14s | 392,559,404 |+| OpenClaw + OpenViking | **82.08%** | 38.8s | 37,423,456 |+| Hermes native memory | 33.38% | 82.4s | 79,228,398 |+| Hermes + OpenViking | **82.86%** | **27.9s** | 52,026,755 |+| Claude Code auto-memory | 57.21% | 49.1s | 353,306,422 |+| Claude Code + OpenViking | **80.32%** | **20.4s** | 129,968,899 |++### 1.2 Key Efficiency Improvements++| Agent | Accuracy Improvement | Latency Reduction | Token Reduction |+| ----- | -------------------- | ----------------- | --------------- |+| OpenClaw | 24.20% → 82.08% (+3.39×) | -59.22% | **-91.0%** |+| Hermes | 33.38% → 82.86% (+2.48×) | -66.10% | -34.3% |+| Claude Code | 57.21% → 80.32% (+1.40×) | -58.45% | -63.2% |++### 2. Agent Experience Memory on tau2-bench++For multi-turn agent tasks on tau2-bench, OpenViking's experience memory improves task success in both retail and airline domains:++| Setting | Retail Accuracy | Airline Accuracy |+| ------- | --------------- | ---------------- |+| LLM without memory | 70.94% | 54.38% |+| LLM + OpenViking experience memory | **77.81%** (+6.87pp) | **66.25%** (+11.87pp) |++### 3. Knowledge Base QA on HotpotQA++On multi-hop RAG tasks from HotpotQA, increasing OpenViking retrieval from top-5 to top-20 delivers the highest accuracy in this comparison while keeping retrieval latency low:++| Method | Retrieval Pattern | Accuracy | Tokens / QA | Latency / QA |+| ------ | ----------------- | -------- | ----------- | ------------ |+| Naive RAG | Vector retrieval | 62.50% | 1,290 | **0.11s** |+| HippoRAG 2 | Vector + knowledge graph | 61.00% | 726 | 20s |+| LightRAG | Vector + knowledge graph | 89.00% | 28,443 | 75s |+| LangChain SQL (Agent) | SQL agent | 78.00% | 4,776 | 132s |+| OpenViking (top-5) | Vector retrieval | 72.75% | 3,154 | 0.22s |+| OpenViking (top-20) | Vector retrieval | **91.00%** | 12,533 | 0.23s |+| Nanobot + OpenViking (Agent) | Vector retrieval + Agent | 87.00% | 71,300 | 61.6s |++### 3.2 Single-turn RAG Across 5 Open-source Datasets++| Method | Retrieval Pattern | Average Accuracy | Indexing Tokens | Tokens / QA | Retrieval Latency |+| ------ | ----------------- | ---------------- | --------------- | ----------- | ----------------- |+| Naive RAG | Vector retrieval | 53.93% | 2,755,356 | 1,435 | **0.13s** |+| PageIndex | Vector + tree structure | 36.75% | 5,609,206 | 710,480 | 84.60s |+| HippoRAG 2 | Vector + knowledge graph | 44.50% | 124,963,618 | **637** | 18.83s |+| LightRAG | Vector + knowledge graph | **76.00%** | 62,705,469 | 27,035 | 9.19s |+| **OpenViking** | **Vector retrieval** | **66.87%** | **8,671,538** | **3,060** | **0.19s** |++> Datasets: FinanceBench, NaturalQuestions, ClapNQ, Qasper, and SyllabusQA. OpenViking reaches 66.87% average accuracy with very low retrieval latency (0.19s), while indexing cost is only 13.8% of LightRAG.+## Quick Start### Prerequisites@@ -605,37 +665,6 @@ To ensure optimal storage performance and data security, we recommend deploying👉 **[View: Server Deployment & ECS Setup Guide](./docs/en/getting-started/03-quickstart-server.md)**-## OpenClaw Context Plugin Details--- Test Dataset: Effect testing based on LoCoMo10 (<https://github.com/snap-research/locomo>) long-range dialogues (1,540 cases in total after removing category5 without ground truth)-- Experimental Groups: Since users may not disable OpenClaw's native memory when using OpenViking, we added experimental groups with native memory enabled or disabled-- OpenViking Version: 0.1.18-- Model: seed-2.0-code-- Evaluation Script: <https://github.com/ZaynJarvis/openclaw-eval/tree/main>--| Experimental Group | Task Completion Rate | Cost: Input Tokens (Total) |-| ------------------------------------------- | -------------------- | -------------------------- |-| OpenClaw(memory-core) | 35.65% | 24,611,530 |-| OpenClaw + LanceDB (-memory-core) | 44.55% | 51,574,530 |-| OpenClaw + OpenViking Plugin (-memory-core) | 52.08% | 4,264,396 |-| OpenClaw + OpenViking Plugin (+memory-core) | 51.23% | 2,099,622 |--- Experimental Conclusions:- After integrating OpenViking:--* With native memory enabled: 43% improvement over original OpenClaw with 91% reduction in input token cost; 15% improvement over LanceDB with 96% reduction in input token cost.-* With native memory disabled: 49% improvement over original OpenClaw with 83% reduction in input token cost; 17% improvement over LanceDB with 92% reduction in input token cost.--👉 **[View: OpenClaw Context Plugin](examples/openclaw-plugin/README.md)**--👉 **[View: OpenCode Memory Plugin Example](examples/opencode-memory-plugin/README.md)**--👉 **[View: Claude Code Memory Plugin Example](examples/claude-code-memory-plugin/README.md)**--👉 **[View: LangChain / LangGraph Integration](./docs/en/agent-integrations/07-langchain-langgraph.md)**--\---## Core ConceptsAfter running the first example, let's dive into the design philosophy of OpenViking. These five core concepts correspond one-to-one with the solutions mentioned earlier, together building a complete context management system:diff --git a/README_CN.md b/README_CN.mdindex ce8eae10f..05f4e8cc0 100644--- a/README_CN.md+++ b/README_CN.md@@ -21,7 +21,7 @@👋 加入我们的社区-📱 <a href="./docs/en/about/01-about-us.md#lark-group">飞书群</a> · <a href="./docs/en/about/01-about-us.md#wechat-group">微信群</a> · <a href="https://discord.com/invite/eHvx8E9XF3">Discord</a> · <a href="https://x.com/openvikingai">X</a>+📱 <a href="./docs/zh/about/01-about-us.md#lark-group">飞书群</a> · <a href="./docs/zh/about/01-about-us.md#wechat-group">微信群</a> · <a href="https://discord.com/invite/eHvx8E9XF3">Discord</a> · <a href="https://x.com/openvikingai">X</a></div>@@ -276,7 +276,7 @@ openviking-server doctor"provider" : "<provider-type>", // 提供商类型:"volcengine"、"openai"、"azure" 等"api_version": "2025-01-01-preview", // (仅 azure)API 版本,可选,默认 "2025-01-01-preview""dimension": 1024, // 向量维度- "model" : "<model-name>" // Embedding 模型名称或 Azure 部署名(如 doubao-embedding-vision-251215 或 text-embedding-3-large)+ "model" : "<model-name>" // Embedding 模型名称或 Azure 部署名},"max_concurrent": 10 // 最大并发 embedding 请求(默认:10)},@@ -285,7 +285,7 @@ openviking-server doctor"api_key" : "<your-api-key>", // 模型服务 API Key(openai-codex 可选)"provider" : "<provider-type>", // 提供商类型 (volcengine, openai, azure, openai-codex 等)"api_version": "2025-01-01-preview", // (仅 azure)API 版本,可选,默认 "2025-01-01-preview"- "model" : "<model-name>", // VLM 模型名称或 Azure 部署名(如 doubao-seed-2-0-pro-260215 或 gpt-4-vision-preview)+ "model" : "<model-name>", // VLM 模型名称或 Azure 部署名"max_concurrent": 100 // 语义处理的最大并发 LLM 调用(默认:100)}}@@ -307,7 +307,7 @@ openviking-server doctor},"log": {"level": "INFO",- "output": "stdout" // 日志输出:"stdout" 或 "file"+ "output": "stdout"},"embedding": {"dense": {@@ -341,7 +341,7 @@ openviking-server doctor},"log": {"level": "INFO",- "output": "stdout" // 日志输出:"stdout" 或 "file"+ "output": "stdout"},"embedding": {"dense": {@@ -533,32 +533,104 @@ ov chat👉 **[查看:服务器部署与 ECS 设置指南](./docs/zh/getting-started/03-quickstart-server.md)**+---++## 📊 评测结果++OpenViking 的核心价值主张:**在更高问答准确率的同时,消耗更低的 Token,完成任务时延更低**。以下结果覆盖三个评测维度。++### 1. 用户记忆评测(User Memory)++**测试目标**:验证 OpenViking 作为不同 Agent 的外接记忆组件,在长对话记忆问答(LOCOMO 数据集)上的准确率、Token 效率和时延表现。++#### 1.1 各 Agent 基座上的 LOCOMO 测试结果++| 实验编号 | 方案 | Query 平均耗时 | 问答准确率 | Agent 总输入 Token |+|---------|------|--------------|-----------|------------------|+| **OpenClaw 基座** | | | | |+| 1 | OpenClaw + 原生 memory-core | 95.14s | 24.20% | 392,559,404 |+| 2 | OpenClaw + Mem0 | **37.6s** | 56.62% | 42,118,285 |+| 3 | OpenClaw + SuperMemory | 109.3s | 42.99% | 88,304,113 |+| 4 | OpenClaw + 百炼记忆库 | 41.6s | 39.55% | 35,206,037 |+| **5** | **OpenClaw + OpenViking** | **38.8s** | **82.08%** | **37,423,456** |+| **Hermes 基座** | | | | |+| 6 | Hermes Native Memory | 82.4s (3.57 轮/query) | 33.38% | 79,228,398 |+| **7** | **Hermes + OpenViking** | **27.9s** (1.55 轮/query) | **82.86%** | **52,026,755** |+| **Claude Code 基座** | | | | |+| 8 | Claude Code Auto-Memory | 49.1s (7.2 轮/query) | 57.21% | 353,306,422 |+| **9** | **Claude Code + OpenViking** | **20.4s** (2.6 轮/query) | **80.32%** | **129,968,899** |++#### 1.2 关键效率提升汇总++| Agent | 准确率提升 | 时延降低 | Token 消耗降低 |+|------|-----------|---------|--------------|+| OpenClaw | 24.20% → 82.08% (+3.39×) | -59.22% | **-91.0%** |+| Hermes | 33.38% → 82.86% (+2.48×) | -66.10% | -34.3% |+| Claude Code | 57.21% → 80.32% (+1.40×) | -58.45% | -63.2% |++---++### 2. Agent 经验记忆评测(Agent Memory)++**测试目标**:验证 OpenViking 抽取并召回经验记忆前后,任务执行表现和 Token 节省的效果。++OpenViking 的 Agent Memory 分为两层:++| 层级 | 概念 | 说明 |+|-----|------|------|+| Layer 1 | **轨迹(Trajectory)** | 每次会话结束后自动提炼,记录"做了什么、怎么做的、结果如何" |+| Layer 2 | **经验(Experience)** | 由多条相关轨迹归纳而来,跨会话可复用的策略性知识,"Situation / Approach / Reflect"三段式 |-## OpenClaw 上下文插件详情+#### 2.1 经济仿真测试(ClawWork)-* 测试集:基于 LoCoMo10(https://github.com/snap-research/locomo) 的长程对话进行效果测试(去除无真值的 category5 后,共 1540 条 case)-* 实验组:因用户在使用 OpenViking 时可能不关闭 OpenClaw 原生记忆,所以增加是否开关原生记忆的实验组-* OpenViking 版本:0.1.18-* 模型:seed-2.0-code-* 评测脚本:https://github.com/ZaynJarvis/openclaw-eval/tree/main+港大数据科学实验室(HKUDS)构建的"实时经济生存 benchmark",Agent 从 $10 起步,每次 LLM 调用自动扣费,收入来自完成专业任务(覆盖 44 个职业、220 个任务)。-| 实验组 | 任务完成率 | 成本:输入 token (总计) |-|----------|------------------|------------------|-| OpenClaw(memory-core) | 35.65% | 24,611,530 |-| OpenClaw + LanceDB (-memory-core) | 44.55% | 51,574,530 |-| OpenClaw + OpenViking Plugin (-memory-core) | 52.08% | 4,264,396 |-| OpenClaw + OpenViking Plugin (+memory-core) | 51.23% | 2,099,622 |+| 方案 | 完成 50 个任务后净收入 | 平均每小时 Token 消耗 |+|-----|-------------------|------------------|+| LLM only | $2,269.77 | 1,030.3K/h |+| **LLM + OpenViking** | **$3,843.74 (+69.34%)** | **872.4K/h (-22.8%)** |-* 实验结论:-结合 OpenViking 后,若仍开启原生记忆,效果在原 OpenClaw 上提升 43%,输入 token 成本降低 91%;在 LanceDB 上效果提升 15%,输入 token 降低 96%。若关闭原生记忆,效果在原 OpenClaw 上提升 49%,输入 token 成本降低 83%;在 LanceDB 上效果提升 17%,输入 token 降低 92%。+#### 2.2 Tau-2 对话 Agent 测试++Sierra AI 发布的对话式 Agent 评测基准,覆盖 Retail 和 Airline 两个领域。++| 方案 | Retail 正确率 | Airline 正确率 |+|-----|-------------|--------------|+| LLM 无记忆 | 70.94% | 54.38% |+| **LLM + OpenViking 经验记忆** | **77.81% (+6.87pp)** | **66.25% (+11.87pp)** |++----👉 **[查看:OpenClaw 上下文插件](examples/openclaw-plugin/README_CN.md)**+### 3. 知识库问答评测(Knowledge Base QA)-👉 **[查看:OpenCode 记忆插件示例](examples/opencode-memory-plugin/README_CN.md)**+**测试目标**:对比 OpenViking 与其他 RAG 方案在开源 benchmark 上的准确率、Token 效率和时延表现。-👉 **[查看:Claude Code 记忆插件示例](examples/claude-code-memory-plugin/README_CN.md)**+#### 3.1 多跳、多路 RAG 测试(HotpotQA 数据集)-👉 **[查看:LangChain / LangGraph 集成](./docs/zh/agent-integrations/07-langchain-langgraph.md)**+| 方案 | 检索范式 | Accuracy | 每 QA Token | 每 QA 耗时 |+|-----|---------|---------|------------|---------|+| Naive RAG | 向量检索 | 62.50% | 1,290 | **0.11s** |+| HippoRAG 2 | 向量 + 知识图谱 | 61.00% | 726 | 20s |+| LightRAG | 向量 + 知识图谱 | 89.00% | 28,443 | 75s |+| LangChain SQL (Agent) | SQL + Agent | 78.00% | 4,776 | 132s |+| OpenViking (top5) | 向量检索 | 72.75% | 3,154 | 0.22s |+| **OpenViking (top20)** | **向量检索** | **91.00%** | **12,533** | **0.23s** |++> 💡 **在本组对比中,OpenViking 在 HotpotQA top20 配置下取得最高准确率(91%);同时检索延迟仅 0.23s,Token 消耗不到 LightRAG 的一半。**++#### 3.2 单轮 RAG 测试(5 个开源数据集均值)++| 方案 | 检索范式 | 平均 Accuracy | 建库 Token | 每 QA Token | 检索耗时 |+|-----|---------|-------------|----------|------------|--------|+| Naive RAG | 向量检索 | 53.93% | 2,755,356 | 1,435 | **0.13s** |+| PageIndex | 向量 + 树结构 | 36.75% | 5,609,206 | 710,480 | 84.60s |+| HippoRAG 2 | 向量 + 知识图谱 | 44.50% | 124,963,618 | **637** | 18.83s |+| LightRAG | 向量 + 知识图谱 | **76.00%** | 62,705,469 | 27,035 | 9.19s |+| **OpenViking** | **向量检索** | **66.87%** | **8,671,538** | **3,060** | **0.19s** |++> 测试数据集:FinanceBench、NaturalQuestions、ClapNQ、Qasper、SyllabusQA。OpenViking 以极低耗时(0.19s)取得 66.87% 的平均准确率,建库成本仅为 LightRAG 的 13.8%。++---## VikingBot 部署详情@@ -566,7 +638,7 @@ OpenViking 有一个类似 nanobot 的机器人用于交互工作,现已可用👉 **[查看:使用 VikingBot 部署服务器](bot/README_CN.md)**---+---## 核心概念@@ -674,8 +746,8 @@ OpenViking 仍处于早期阶段,有许多改进和探索的空间。我们真- 为我们点亮一颗珍贵的 **Star**,给我们前进的动力。- 访问我们的[**官网**](https://www.openviking.ai)了解我们传达的理念,并通过[**文档**](https://www.openviking.ai/docs)在您的项目中使用它。感受它带来的变化,并给我们最真实的体验反馈。- 加入我们的社区,分享您的见解,帮助回答他人的问题,共同创造开放互助的技术氛围:- - 📱 **飞书群**:扫码加入 → [查看二维码](./docs/en/about/01-about-us.md#lark-group)- - 💬 **微信群**:扫码添加助手 → [查看二维码](./docs/en/about/01-about-us.md#wechat-group)+ - 📱 **飞书群**:扫码加入 → [查看二维码](./docs/zh/about/01-about-us.md#lark-group)+ - 💬 **微信群**:扫码添加助手 → [查看二维码](./docs/zh/about/01-about-us.md#wechat-group)- 🎮 **Discord**:[加入 Discord 服务器](https://discord.com/invite/eHvx8E9XF3)- 🐦 **X (Twitter)**:[关注我们](https://x.com/openvikingai)- 成为**贡献者**,无论是提交错误修复还是贡献新功能,您的每一行代码都将是 OpenViking 成长的重要基石。@@ -708,10 +780,10 @@ OpenViking 项目不同组件采用不同的开源协议:[license-shield]: https://img.shields.io/badge/license-AGPLv3-white?labelColor=black&style=flat-square[license-shield-link]: https://github.com/volcengine/OpenViking/blob/main/LICENSE[last-commit-shield]: https://img.shields.io/github/last-commit/volcengine/OpenViking?color=c4f042&labelColor=black&style=flat-square-[last-commit-shield-link]: https://github.com/volcengine/OpenViking/commcommits/main+[last-commit-shield-link]: https://github.com/volcengine/OpenViking/commits/main[github-stars-shield]: https://img.shields.io/github/stars/volcengine/OpenViking?labelColor&style=flat-square&color=ffcb47[github-stars-link]: https://github.com/volcengine/OpenViking[github-issues-shield]: https://img.shields.io/github/issues/volcengine/OpenViking?labelColor=black&style=flat-square&color=ff80eb[github-issues-shield-link]: https://github.com/volcengine/OpenViking/issues[github-contributors-shield]: https://img.shields.io/github/contributors/volcengine/OpenViking?color=c4f042&labelColor=black&style=flat-square-[github-contributors-link]: https://github.com/volcengine/OpenViking/graphs/contributors+[github-contributors-link]: https://github.com/volcengine/OpenViking/graphs/contributors\ No newline at end of filediff --git a/README_JA.md b/README_JA.mdindex 703332c93..cb00da7f3 100644--- a/README_JA.md+++ b/README_JA.md@@ -57,6 +57,63 @@ OpenVikingを使えば、開発者はローカルファイルを管理するよ---+## 評価ハイライト++OpenVikingの中核的な価値は、**より高い回答精度**を実現しながら、**より少ないトークン消費**と**より低いレイテンシ**を両立することです。以下では、3つの評価シナリオをまとめます。++### 1. ユーザーメモリ評価++**目的**: OpenVikingを各種Agentの外部メモリとして接続したとき、長対話メモリQA(LOCOMO)における精度・Token効率・レイテンシを検証します。++| 構成 | 平均 Query 時間 | 回答精度 | 総入力 Token |+|------|----------------|---------|-------------|+| OpenClaw + native memory-core | 95.14s | 24.20% | 392,559,404 |+| **OpenClaw + OpenViking** | **38.8s** | **82.08%** | **37,423,456** |+| Hermes Native Memory | 82.4s | 33.38% | 79,228,398 |+| **Hermes + OpenViking** | **27.9s** | **82.86%** | **52,026,755** |+| Claude Code Auto-Memory | 49.1s | 57.21% | 353,306,422 |+| **Claude Code + OpenViking** | **20.4s** | **80.32%** | **129,968,899** |++### 1.2 主要な効率改善++| Agent | 精度向上 | レイテンシ低減 | Token 削減 |+|------|---------|---------------|-----------|+| OpenClaw | 24.20% → 82.08% (+3.39×) | -59.22% | **-91.0%** |+| Hermes | 33.38% → 82.86% (+2.48×) | -66.10% | -34.3% |+| Claude Code | 57.21% → 80.32% (+1.40×) | -58.45% | -63.2% |++### 2. Agent経験メモリ評価++| 構成 | Retail 正答率 | Airline 正答率 |+|------|--------------|---------------|+| LLMのみ | 70.94% | 54.38% |+| **LLM + OpenViking 経験メモリ** | **77.81% (+6.87pp)** | **66.25% (+11.87pp)** |++### 3. ナレッジベースQA評価++| 手法 | 検索方式 | Accuracy | Token / QA | レイテンシ / QA |+|------|---------|----------|------------|-----------------|+| Naive RAG | ベクトル検索 | 62.50% | 1,290 | **0.11s** |+| HippoRAG 2 | ベクトル + 知識グラフ | 61.00% | 726 | 20s |+| LightRAG | ベクトル + 知識グラフ | 89.00% | 28,443 | 75s |+| LangChain SQL (Agent) | SQL + Agent | 78.00% | 4,776 | 132s |+| OpenViking (top5) | ベクトル検索 | 72.75% | 3,154 | 0.22s |+| **OpenViking (top20)** | **ベクトル検索** | **91.00%** | **12,533** | **0.23s** |++> この比較では、OpenVikingは HotpotQA の top20 設定で最高精度(91%)を達成し、同時にレイテンシを 0.23s に抑えています。++### 3.2 5つのオープンソースデータセットにおける単一ターンRAG++| 手法 | 検索方式 | 平均 Accuracy | インデックス構築 Token | Token / QA | 検索レイテンシ |+|------|---------|---------------|------------------------|------------|----------------|+| Naive RAG | ベクトル検索 | 53.93% | 2,755,356 | 1,435 | **0.13s** |+| PageIndex | ベクトル + 木構造 | 36.75% | 5,609,206 | 710,480 | 84.60s |+| HippoRAG 2 | ベクトル + 知識グラフ | 44.50% | 124,963,618 | **637** | 18.83s |+| LightRAG | ベクトル + 知識グラフ | **76.00%** | 62,705,469 | 27,035 | 9.19s |+| **OpenViking** | **ベクトル検索** | **66.87%** | **8,671,538** | **3,060** | **0.19s** |++> 対象データセット: FinanceBench、NaturalQuestions、ClapNQ、Qasper、SyllabusQA。OpenViking は非常に低い検索レイテンシ(0.19s)を維持しながら平均 66.87% の精度を達成し、インデックス構築コストは LightRAG の 13.8% にとどまります。+## クイックスタート### 前提条件@@ -396,34 +453,6 @@ ov chat👉 **[参照: サーバーデプロイ&ECSセットアップガイド](./docs/en/getting-started/03-quickstart-server.md)**-## OpenClawコンテキストプラグインの詳細--* テストデータセット: LoCoMo10(https://github.com/snap-research/locomo)の長距離対話に基づく効果テスト(ground truthのないcategory5を除いた合計1,540ケース)-* 実験グループ: ユーザーがOpenVikingを使用する際にOpenClawのネイティブメモリを無効にしない可能性があるため、ネイティブメモリの有効/無効の実験グループを追加-* OpenVikingバージョン: 0.1.18-* モデル: seed-2.0-code-* 評価スクリプト: https://github.com/ZaynJarvis/openclaw-eval/tree/main--| 実験グループ | タスク完了率 | コスト: 入力トークン数(合計) |-|----------|------------------|------------------|-| OpenClaw(memory-core) | 35.65% | 24,611,530 |-| OpenClaw + LanceDB (-memory-core) | 44.55% | 51,574,530 |-| OpenClaw + OpenViking Plugin (-memory-core) | 52.08% | 4,264,396 |-| OpenClaw + OpenViking Plugin (+memory-core) | 51.23% | 2,099,622 |--* 実験結果:-OpenViking統合後:-- ネイティブメモリ有効時: オリジナルOpenClawと比較して43%改善、入力トークンコスト91%削減。LanceDBと比較して15%改善、入力トークンコスト96%削減。-- ネイティブメモリ無効時: オリジナルOpenClawと比較して49%改善、入力トークンコスト83%削減。LanceDBと比較して17%改善、入力トークンコスト92%削減。--👉 **[参照: OpenClawコンテキストプラグイン](examples/openclaw-plugin/README.md)**--👉 **[参照: OpenCodeメモリプラグインの例](examples/opencode-memory-plugin/README.md)**--👉 **[参照: Claude Codeメモリプラグインの例](examples/claude-code-memory-plugin/README.md)**-----## コアコンセプト最初の例を実行した後、OpenVikingの設計思想を掘り下げましょう。これら5つのコアコンセプトは、先に述べたソリューションと1対1で対応し、完全なコンテキスト管理システムを構築します: