From 152dc680ad7460201d5b505f0a7d357aa94c532f Mon Sep 17 00:00:00 2001 From: JianFeeeee Date: Sat, 31 Oct 2026 11:30:18 +0800 Subject: [PATCH] =?UTF-8?q?docs:=20=E6=9E=B6=E6=9E=84=E8=8A=82=E5=86=99?= =?UTF-8?q?=E6=B8=85=E4=B8=8A=E4=B8=8B=E6=96=87=E7=AE=A1=E7=90=86=E4=B8=8E?= =?UTF-8?q?=E8=AE=B0=E5=BF=86=E7=9A=84=E5=AE=9E=E9=99=85=E5=81=9A=E6=B3=95?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ## 补上真正的差异点 之前架构节只有一张图和七阶段表，把最核心的上下文管理省略了。 用 codegraph 查了 `RelevanceContext` / `PruneWithProtected` / `coldcontext.Store` / `block_recall`，确认实际做法，逐条对照常见 agent： ### 上下文：按相关度取舍，不是截断或摘要 常见做法是超窗裁掉旧对话，或让模型自己摘要 —— 共同问题是**丢什么由位置决定**。 这里是给每条历史事件按当前输入打分，低分下沉到文档层。 打分主路径是稠密余弦（与媒体、文档同一向量空间），模型未就绪退到 TF-IDF。 关键在 `effectiveProtectedCount`：保护条数受预算约束。 ```go budgetCap = ContextTokens / avgEventTokens pCount = min(配置值, budgetCap) ``` 源码里记着这条约束的来历：50k 窗口下单条事件约 5800 token（工具回灌型）， 预算 40000，「钉住 10 条」本身就装不下 —— 每轮裁剪都撞保护下限， 两轮后预算耗尽，任务直接终止，召回率从 44% 掉到 22%。 ### 三层是迁移，不是复制 `ColdDoc` 的注释写得很清楚：「块的真实归属在 Events[i].Blocks 里」， 块在任一层只有一个持有者，不靠引用计数保活。 回灌用 `Peek` + 确认 + `Remove` 两阶段，源码明确标注「**禁止先删后灌** —— 删除必须排在确认之后，回灌失败即永久丢数据」。 ### 场面：两条路并行，且互不吞并 声明场景**不吸收整轮指纹**（`EnsureScene` 的注释）：否则它会以接近 1.0 的相似度 吃掉后续所有同类轮次，被动聚类再也长不出更细的场面。 这条边界是踩出来的，不是设计时想到的。 ## 开头相应收缩 原来开头四段「**粗体断言** + 三行解释」与架构节重复。改成一句介绍 + 安装 + 三段话点出差异，细节留给架构节。 两版 163 / 191 行，9 个标题对齐，代码块 10/10 配对，无死链。 --- README.md | 76 +++++++++++++++++++++++++++++----------- README_EN.md | 98 ++++++++++++++++++++++++++++++++++++---------------- 2 files changed, 123 insertions(+), 51 deletions(-) diff --git a/README.md b/README.md index 69bd8b47..16055308 100644 --- a/README.md +++ b/README.md @@ -13,22 +13,9 @@ make build build-cli API 密钥在 WebUI（`http://localhost:8080`）里填。需要 Linux。 -## 几个设计 - -**启动顺序在编译期定序。** 20 个内核模块各自声明 `Requires`/`Provides`，构建时算出 6 层 -依赖图写进 `depgraph.bin`，digest 用 `-ldflags` 注入二进制。运行期只读图，不重算。 -图与二进制不配套则拒绝启动。 - -**插件按设施就绪提前启动。** 声明 `requires_facilities: [config]` 的插件在第 1 层完成时就起， -不必等第 4 层 agent。实测省 35 秒，那是 agent 层做 dense backfill 750 条事件、重建索引的时间。 - -**两级输入、四级中断。** 排队输入无级别，任何中断可打断。中断带 L1–L4，高抢占低并将现场 -压栈待恢复，同级不抢占。L4 内核独占，插件声明会被夹到 L3。抢占计数可提级，上限两档 -（`preemptPromotionCap = 2`），避免低优先级任务长期得不到执行。 - -**记忆按场面分。** 指纹由通道、对话对象、工具、话题、时段构成，权重 1.0/1.0/0.8/0.4/0.2。 -同一指纹出现到第 2 次才建场景（`minSceneEvidence = 2`）。归属阈值 0.5、唤起阈值 0.35， -唤起比归属宽松：想不起来是损失，多想一条只是多几行上下文。 +跟常见 agent 框架的主要差别在上下文管理：不是超窗就截断或让模型摘要， +而是按相关度取舍，低分的下沉成文档、再蒸馏进图数据库。另外两处不太一样的是 +启动顺序在编译期定序，以及插件按设施就绪提前启动（详见下方架构）。 ## 架构 @@ -39,7 +26,7 @@ API 密钥在 WebUI（`http://localhost:8080`）里填。需要 Linux。 ┌──────────────┼──────────────┐ │ │ │ stdio RPC 共享内存 事件环 - 控制 大块数据 通知 + 控制 大块数据 通知 │ │ │ └──────────────┼──────────────┘ │ @@ -50,7 +37,57 @@ API 密钥在 WebUI（`http://localhost:8080`）里填。需要 Linux。 内核不发 HTTP、不读文件。所有接触外部世界的能力都在插件里，插件崩溃不影响内核， 重启退避 1s / 2s / 3s，5 分钟内第 4 次崩溃停止自动重启等人工介入。 -**消息进入后的七个阶段**（`internal/agent/core/stages.go`）： +### 上下文不是「截断 + 摘要」 + +常见做法是把超窗的对话裁掉或让模型自己摘要，两者的共同问题是**丢什么由位置决定**。 +这里换成了按相关度取舍：用当前输入给每条历史事件打分，低分的下沉到文档层。 + +打分用稠密向量余弦（与媒体、文档共享同一向量空间），模型未就绪时退化到 TF-IDF。 +裁剪时无条件保住最近 N 条，但 N 本身受预算约束： + +```go +// 保护条数不超过预算能装下的条数 +budgetCap = ContextTokens / avgEventTokens +pCount = min(配置值, budgetCap) +``` + +这条约束来自一次实测：50k 窗口下单条事件约 5800 token（工具回灌型），预算约 40000， +「钉住 10 条」本身就装不下，于是每轮裁剪都撞保护下限，两轮后预算耗尽，任务直接终止。 +召回率从 44% 掉到 22%。现在窗口越小越自动收紧，1M 窗口下则能用满配置值。 + +被裁掉的事件不是丢弃，是迁移：事件连同它持有的媒体块一起进文档层，块的 ID 不变、 +只换持有层。回灌时用 `coldcontext_query` 取回，两阶段（先 Peek 确认、再 Remove）， +不用「先删后灌」——那会在回灌失败时永久丢数据。 + +### 三层记忆是迁移，不是复制 + +``` +L0 上下文 相关性裁剪，最近 N 条恒受保护 + ↓ 低分事件下沉（带媒体块） +L1 文档 稠密检索，访问计数 ≤2 且久未访问则继续下沉 + ↓ 蒸馏成三元组 +L2 图 实体关系 + 场面索引，支持按场面召回 +``` + +块在任一层只有一份持有者，不靠引用计数保活。媒体不是附件：图片、音频是图里的一等节点， +有自己的向量，和文本在同一空间里检索。 + +### 记忆按场面分，不按会话 + +普通 agent 通常一条会话一条历史，换个话题就失忆，或者全塞进一个上下文里互相干扰。 +这里按「可观察的场所指纹」分：通道、对话对象、工具、话题、时段，权重 1.0/1.0/0.8/0.4/0.2。 + +同一指纹出现到第 2 次才建场景（`minSceneEvidence = 2`），只发生一次的不算。 +归属阈值 0.5、唤起阈值 0.35，唤起比归属宽松——想不起来是损失，多想一条只是多几行上下文。 +场景每次重现强度 +1，久不重现按半衰期衰减。 + +两个来源并行：**声明**（插件在注入点说明"这是哪个场面"）与**涌现**（指纹自己聚类）。 +声明场景不吸收整轮指纹，否则它会以接近 1.0 的相似度吃掉后续所有同类轮次， +被动路再也长不出更细的场面。 + +### 七个阶段 + +`internal/agent/core/stages.go` | 阶段 | 时机 | 插件可做 | |---|---|---| @@ -65,9 +102,6 @@ API 密钥在 WebUI（`http://localhost:8080`）里填。需要 Linux。 `post_action → [before_toolcall → 执行 → after_toolcall] → post_action` 是内循环， 退出条件：模型不再调工具、全部被拒、或达到上限。 -**三层记忆**：上下文（最近 10 条恒受保护）→ 文档（冷数据下沉）→ 图数据库（蒸馏成三元组）。 -媒体是图里的一等节点，不是附件。 - ## 其余 - **通道**：WebUI、命令行、QQ 群与私聊、A2A、IDE 里的 ACP diff --git a/README_EN.md b/README_EN.md index a21fcbc3..cdcb5db1 100644 --- a/README_EN.md +++ b/README_EN.md @@ -13,28 +13,10 @@ make build build-cli API keys go in the web UI at `http://localhost:8080`. Requires Linux. -## Design notes - -**Startup order is resolved at build time.** Twenty kernel modules declare `Requires`/`Provides`. -The build computes a 6-layer graph into `depgraph.bin` and injects its digest into the binary via -`-ldflags`. At runtime the kernel reads the graph rather than recomputing it. A mismatch between -graph and binary refuses startup. - -**Plugins launch as soon as their facilities are ready.** A plugin declaring -`requires_facilities: [config]` starts when layer 1 completes instead of waiting for the agent -layer at layer 4. That saves a measured 35 seconds, which the agent layer spends on dense -backfill of 750 events and reindexing. - -**Two input classes, four interrupt levels.** Queued input has no level, so any interrupt can -preempt it. Interrupts carry L1–L4: higher preempts lower and pushes the preempted state onto a -stack for later resume, same level never preempts. L4 is kernel-only; plugins declaring it get -clamped to L3. Preemption count can raise a level by at most two steps -(`preemptPromotionCap = 2`), so low-priority work can't starve indefinitely. - -**Memory is partitioned by scene.** A fingerprint is built from channel, partner, tools, topic, -and time of day, weighted 1.0/1.0/0.8/0.4/0.2. A scene exists only after the second matching -fingerprint (`minSceneEvidence = 2`). Join threshold 0.5, recall 0.35 — recall is looser because -failing to recall loses information while an extra scene costs a few lines of context. +The main difference from typical agent frameworks is context handling: instead of truncating +on overflow or asking the model to summarize, it scores history by relevance and sinks the +low-scoring events into documents, then distills those into a graph. Two other things differ: +startup order is resolved at build time, and plugins launch as soon as their facilities are ready. ## Architecture @@ -53,11 +35,70 @@ failing to recall loses information while an extra scene costs a few lines of co web UI · QQ · CLI · files · browser ``` -The kernel makes no HTTP calls and reads no files. Everything that touches the outside world is a -plugin. A plugin crash leaves the kernel running; restart backs off 1s / 2s / 3s, and the 4th -crash within 5 minutes stops automatic restarts for human intervention. +The kernel makes no HTTP calls and reads no files. Everything touching the outside world is a +plugin. A plugin crash leaves the kernel running; restarts back off 1s / 2s / 3s, and the 4th +crash within 5 minutes stops automatic restarting for human intervention. -**Seven stages after a message arrives** (`internal/agent/core/stages.go`): +### Context is not "truncate and summarize" + +The usual approach drops old turns or asks the model to summarize them. Both share a flaw: what +gets lost is decided by position. Here each historical event is scored against the current input, +and low scorers sink into the document layer. + +Scoring uses dense cosine similarity in the same vector space as media and documents, falling +back to TF-IDF when the model isn't loaded. The most recent N events are always kept, but N is +itself bounded by budget: + +```go +// protected count cannot exceed what the budget can hold +budgetCap = ContextTokens / avgEventTokens +pCount = min(configured, budgetCap) +``` + +That bound came from a measurement: with a 50k window, one event averaged 5800 tokens (tool +results), while the budget was 40000, so "pin 10 events" didn't fit. Every prune hit the +protection floor, the budget was exhausted after two rounds, and the task terminated. Recall +dropped from 44% to 22%. Smaller windows now tighten automatically, and a 1M window can use the +full configured value. + +Pruned events aren't discarded, they migrate: the event and the media blocks it holds move into +the document layer with block IDs unchanged, only the holding layer changes. Retrieval uses +`coldcontext_query` in two phases (Peek to confirm, then Remove) rather than delete-then-refill, +which would lose data permanently if the refill failed. + +### Three layers migrate, they don't copy + +``` +L0 context relevance pruning, last N always protected + ↓ low scorers sink (with media blocks) +L1 documents dense retrieval; access count ≤2 and stale means sink further + ↓ distilled into triples +L2 graph entities, relations, scene index; recall by scene +``` + +A block has exactly one holder at any layer, with no reference counting to keep it alive. Media +are not attachments: images and audio are first-class graph nodes with their own vectors, +searchable in the same space as text. + +### Memory is partitioned by scene, not by session + +Typical agents keep one history per session, so switching topics loses context or crams +everything into one window. Here memory is partitioned by observable scene fingerprint: channel, +conversation partner, tools, topic, time of day, weighted 1.0/1.0/0.8/0.4/0.2. + +A scene comes into being only on the second matching fingerprint (`minSceneEvidence = 2`). +Join threshold 0.5, recall 0.35 — recall is looser because failing to recall loses information +while an extra scene costs a few lines of context. Each recurrence adds to scene strength; +scenes that stop recurring decay by half-life. + +Two paths run in parallel: **declared** (a plugin states at the injection point which scene this +is) and **emergent** (fingerprints cluster on their own). Declared scenes deliberately don't +absorb the full fingerprint, otherwise they'd score near 1.0 against every later turn and the +emergent path could never grow finer-grained scenes. + +### Seven stages + +`internal/agent/core/stages.go` | Stage | When | Plugins can | |---|---|---| @@ -67,14 +108,11 @@ crash within 5 minutes stops automatic restarts for human intervention. | `before_toolcall` | before one tool runs | audit, deny, rewrite args | | `after_toolcall` | after one tool runs | redact, sort results | | `before_output` | final text ready | format adaptation | -| `after_output` | already sent | statistics | +| `after_output` | after sending | statistics | `post_action → [before_toolcall → run → after_toolcall] → post_action` is the inner loop. It exits when the model stops calling tools, all calls are denied, or the limit is reached. -**Three memory layers**: context (last 10 always protected) → documents (cold data sinks) → graph -database (distilled into triples). Media are first-class graph nodes, not attachments. - ## Everything else - **Channels**: web UI, CLI, QQ groups and DMs, A2A, ACP in your IDE