Belief Context Graph：把 Agent 记忆从「检索」升级为「置信度感知的信念图」

现有 Agent 记忆的盲区

当前主流 Agent 记忆系统做的事情本质上都是同一件：检索。对话记忆保存历史，向量记忆找相似片段，GraphRAG 提取实体关系，Trace 记忆记录工具调用……

它们回答的都是”检索问题”：哪段文字相关？哪些实体有关联？过去发生了什么？

但执行真实任务的 Agent 还需要回答另一类问题——信念问题：

• 我应该相信这个事实吗？
• 它还有效，还是已经过期？
• 来源可靠吗？
• 它和其他证据矛盾吗？
• 确定性够不够，值得采取行动吗？
• 结果证明我之前的判断是错的吗？

Belief Context Graph（BCG） 是北京通用人工智能研究院（BIGAI）开源的一个 Agent 记忆基础设施，专门回答这类问题。

---

BCG 的核心：四个能力

1. 信念提取（Belief Extraction）

BCG 把 Agent 的执行轨迹切分成片段，从中提取结构化信念节点，每个节点记录一个命题（“X 是 Y”、“A 导致 B”……），并把这些节点连接成图。

这和 GraphRAG 的实体提取不同：BCG 提取的是 Agent 推理过程中实际依赖的命题，而不是文本中出现的所有实体。

2. 可审计的确定性置信度（Deterministic Confidence）

这是 BCG 与其他系统最大的区别。每个信念节点的置信度由三个组件确定性计算得出：

posterior_confidence = f(
    initial_confidence,   // 初始置信度（来源可靠性 + 立场质量）
    evidence_confidence,  // 证据置信度（支持 / 反对证据的累积）
    factor_confidence     // 关系推导置信度（相邻节点的传播权重）
)

这个计算是确定性的（deterministic），不是 LLM 给出的模糊评分。你可以追溯任何一个信念节点的置信度是怎么算出来的，哪条证据贡献了多少权重。

3. 证据溯源（Evidence Provenance）

每个信念节点携带精确的来源引用——具体是哪一轮对话的哪个偏移量产生了这个信念。Agent 可以知道”这个结论来自第 3 轮工具调用返回的第二段文字”。

4. 时间感知与关系链接

• Temporal Awareness：运行级生命周期，记录每个信念形成的时间戳和演化轨迹
• Relation Linking：信念节点之间有前向/后向关系边，形成因果决策图/追踪链

---

架构：Agent 和模型之间的可选上下文层

BCG 作为一个可选的上下文层插在 Agent 和模型之间：

用户输入 + 最近几轮 → [保留在原始上下文]
更早的已完成轮次   → [流入 Graph Construction]
                              ↓
                   Belief Snapshot（信念快照）
                              ↓
                   注入到 system prompt
                              ↓
                    Agent → LLM 推理

较旧的已完成轮次不再堆在上下文里，而是被蒸馏成信念图快照，只有被提取为信念节点的内容（附带置信度和来源）才会出现在 system prompt 里。这同时压缩了 token 用量，并提升了推理质量。

HTTP 服务和 Python SDK 使用同一套后端注册表、构建流水线、置信度语义和图产物。

---

实测案例：BrowseComp 任务

在 BrowseComp 基准测试中，使用 BCG 的 Kimi K3 在任务中途识别出”这个搜索我已经做过了”（通过信念节点的 identity 和置信度），直接跳过重复搜索，而不是重跑一遍。这正是信念图解决”长程 Agent 重复行为”的典型模式。

基准结果：BCG 在准确率和 token 成本两个维度上均有改善（详见 benchmark overview）。

---

与主流方案的对比

	Mem0	Zep	LangChain	LlamaIndex	Semantica	BCG
信念原生提取	⚡	⚡	⚡	⚡	⚡	✅
确定性置信度	❌	❌	❌	❌	⚡	✅
证据溯源	⚡	✅	⚡	⚡	✅	✅
冲突检测	⚡	✅	⚡	❌	✅	✅
本地产物（无外部 DB）	✅	✅	✅	✅	✅	✅
合并/去重	⚡	✅	⚡	⚡	✅	✅

BCG 独有：信念原生提取 + 确定性置信度的组合，在对比表中只有 BCG 同时做到了这两点。

---

快速上手

git clone https://github.com/bigai-nlco/belief-context-graph.git
cd belief-context-graph
make install

# 运行内置参考 Agent（首次运行引导设置模型凭证）
uv run bcg

Python 3.11–3.13，依赖通过 uv 管理。

Python SDK（集成到自己的 Agent）：

from bcg import BCGMemory, BCGRunner

# 构建信念图
runner = BCGRunner(session_id="my-agent-session")
runner.ingest(trajectory)   # 喂入 Agent 执行轨迹

# 查询信念
memory = BCGMemory(session_id="my-agent-session")
beliefs = memory.observe_belief(query="用户的偏好设置")

# 返回：信念节点列表，每个含 content / confidence / evidence_refs / relations

文档：belief-context-graph.docs.buildwithfern.com

---

适合哪些场景

长程 Agent 任务（Deep Research、多步规划、复杂工作流）：对话轮次多，Agent 需要跨步骤引用之前的结论。BCG 防止 Agent 遗忘已建立的结论，也防止它重复执行已完成的步骤。

需要可审计推理的场景（法律/医疗/金融 Agent）：置信度和来源可以被人类检查员追溯——“这个结论置信度 0.73，来自第 5 轮搜索结果的第 2 段，与第 8 轮的新证据有轻微冲突”。

Agent 开发研究：BCG 的信念图和置信度传播机制是一个研究 Agent 推理可靠性的工具，可以可视化 Agent 在任务执行中”信什么、信多少、为什么信”的演化过程。

---

路线图

BCG 下一阶段有两个方向：

1. 更严格的概率基础：把当前的置信度计算迁移到 Bayesian 推断或其他有数学保证的不确定性框架，同时保持可审计性。
2. Deep Research 扩展：从信念感知的上下文管理，扩展为能规划调查路径、追踪来源时效性、调和矛盾发现、识别缺失证据并输出可审计研究报告的完整工作流。

---

总结

BCG 提出了一个简单但重要的区分：Agent 记忆需要回答的不只是”检索问题”（哪些信息相关），还有”信念问题”（这个信息该不该信、还有没有效）。确定性置信度 + 证据溯源 + 冲突检测的组合，是现有主流方案里没有同时做到的。对于构建需要长程推理和可靠性保证的 Agent 系统，BCG 是一个值得认真研究的内存基础设施选项。

GitHub: bigai-nlco/belief-context-graph ⭐59
文档: belief-context-graph.docs.buildwithfern.com
联系: lijiaqi@bigai.ai · zlzheng@bigai.ai

Belief Context Graph: Upgrading Agent Memory from Retrieval to Belief

Current agent memory systems — conversation memory, vector memory, GraphRAG, trace memory — all do the same thing at their core: retrieval. They answer retrieval questions: which text is relevant? which entities are related? what happened before?

But agents executing real tasks also need to answer belief questions:

• Should I actually believe this fact?
• Is it still valid, or has it expired?
• Did it come from a reliable source?
• Does it conflict with other evidence?
• Is it certain enough to act on?

Belief Context Graph (BCG) from BIGAI (Beijing Institute for General Artificial Intelligence) is a memory substrate specifically designed to answer these questions.

Four Core Capabilities

Belief Extraction: BCG segments agent trajectories and extracts structured belief nodes — propositions that the agent’s reasoning actually depends on — and links them into a connected graph. Unlike GraphRAG’s entity extraction, BCG extracts what the agent believes, not what appeared in the text.

Deterministic Confidence: The most distinctive feature. Each belief node’s confidence is computed deterministically from three components:

posterior_confidence = f(
    initial_confidence,   // source reliability + stance quality
    evidence_confidence,  // accumulated supporting/contradicting evidence
    factor_confidence     // propagated weights from related nodes
)

This is deterministic — not a fuzzy LLM score. You can trace exactly how any belief’s confidence was calculated and which evidence contributed what weight.

Evidence Provenance: Every belief node carries exact-offset source references back to the specific conversation turn and position that produced it. The agent can know: “this conclusion came from the second paragraph of the tool call result in turn 3.”

Temporal Awareness + Relation Linking: Run-based lifecycle with timestamps records when each belief formed and how it evolved. Forward and backward relationship edges between belief nodes form a causal decision graph.

Architecture

BCG inserts as an optional context layer between the Agent and the model:

Recent turns          → [stay in raw context]
Older completed turns → [stream into Graph Construction]
                                  ↓
                       Belief Snapshot
                                  ↓
                       Injected into system prompt

Older turns don’t pile up in the context window — they’re distilled into a belief graph snapshot. Only what was extracted as belief nodes (with confidence and provenance) appears in the system prompt, simultaneously compressing token usage and improving reasoning quality.

Benchmark Results

In BrowseComp testing, a Kimi K3 agent using BCG recognized mid-task that a search had already been completed (via belief node identity and confidence) and skipped the redundant search — exactly the pattern BCG addresses: preventing long-horizon agents from repeating completed steps.

Comparison

Among Mem0, Zep, Letta, LangChain Memory, LlamaIndex, TrustGraph, and Semantica — BCG is the only system that combines belief-native extraction + deterministic confidence + conflict detection. Most systems have retrieval; none has all three of these.

Quick Start

git clone https://github.com/bigai-nlco/belief-context-graph.git
cd belief-context-graph
make install
uv run bcg   # reference agent with guided first-run setup

Python SDK:

from bcg import BCGMemory, BCGRunner

runner = BCGRunner(session_id="my-session")
runner.ingest(trajectory)

memory = BCGMemory(session_id="my-session")
beliefs = memory.observe_belief(query="user preferences")
# Returns: belief nodes with content / confidence / evidence_refs / relations

Who It’s For

Long-horizon agent tasks (Deep Research, multi-step planning): BCG prevents agents from forgetting established conclusions and repeating completed work across many turns.

Auditable reasoning scenarios (legal/medical/financial agents): Confidence and provenance are traceable by human reviewers — “confidence 0.73, sourced from turn 5 search result paragraph 2, with minor conflict from turn 8 evidence.”

Agent research: BCG’s belief graph and confidence propagation mechanism visualizes how an agent’s beliefs evolve — what it believes, how much, and why — throughout task execution.

Roadmap

1. More principled probabilistic foundation: migrate confidence computation to Bayesian inference or equivalent mathematically justified uncertainty frameworks
2. Deep Research extension: from belief-aware context management to a full workflow that plans investigations, tracks source provenance and temporal validity, reconciles conflicting findings, and produces auditable research outputs

BCG makes a simple but important distinction: agent memory needs to answer not just retrieval questions but belief questions. Deterministic confidence + evidence provenance + conflict detection is a combination no mainstream alternative currently achieves.

GitHub: bigai-nlco/belief-context-graph ⭐59
Docs: belief-context-graph.docs.buildwithfern.com