volcengine · wujiaming-ai · Sep 23, 2026 · Sep 23, 2026 · Sep 24, 2026 · Sep 24, 2026
diff --git a/docs/content/docs/framework/memory/long-term/index.en.mdx b/docs/content/docs/framework/memory/long-term/index.en.mdx
@@ -31,6 +31,8 @@ ltm = LongTermMemory(backend="viking", app_name="ltm_demo")
 | `index` | `str` | `""` | Index/collection name for storing memories. Falls back to `app_name`, then `default_app`. |
 | `app_name` | `str` | `""` | The owning application name; used for data isolation and as the `index` fallback. |
 | `user_id` | `str` | `""` | **Deprecated**, kept only for backward compatibility. |
+| `recall_strategy` | `str` | `"off"` | Recall judgement: with `decision`, the decision model drops memories unrelated to the request. |
+| `recall_relevance_threshold` | `float` | `0.5` | Recall relevance threshold: memories judged below it are not returned. |
 
 <Callout type="info">
 Vector backends (`local`, `opensearch`, `redis`) embed memories, which requires `pip install "veadk-python[extensions]"` and an embedding model (env prefix `MODEL_EMBEDDING_`, falling back to `MODEL_AGENT_API_KEY`). `viking`, `mem0`, `openviking`, and `tos_context` are managed services and need no local embedding.
@@ -173,6 +175,17 @@ agent = Agent(
 ```
 
 To avoid frequent index re-initialization, VeADK exposes `MIN_MESSAGES_THRESHOLD` and `MIN_TIME_THRESHOLD` env vars to tune the save cadence: by default it saves after 10 accumulated events or a 60-second interval; additionally, when you switch `session_id` and start a new turn, VeADK saves the previous session to long-term memory.
+
+The default is `MEMORY_SAVE_STRATEGY=threshold`, the two thresholds above. With `decision`, the configured decision model decides whether the turn holds something durable (a stated preference, fact, decision, or constraint), thresholded by `MEMORY_SAVE_WORTH_THRESHOLD` (default `0.5`):
+
+| Judgement | Behaviour |
+| --- | --- |
+| Above the threshold | Saved immediately, without waiting for 10 events or 60 seconds |
+| Below the threshold | Skipped, and not saved merely because events accumulated |
+| Unavailable | Falls back to `MIN_MESSAGES_THRESHOLD` / `MIN_TIME_THRESHOLD` |
+
+A skipped turn does not advance the save cursor, so the next accepted judgement writes those events together: nothing is lost. See the [environment variable reference](/references/configuration/environment-variables) for the decision model variables.
+Recall can be judged the same way: with `MEMORY_RECALL_STRATEGY=decision`, `search_memory` rates every returned memory (irrelevant / related but useless / useful background / required) and drops the ones below `MEMORY_RECALL_RELEVANCE_THRESHOLD` (default `0.5`). One judgement covers at most 20 memories, anything beyond that is returned in backend order, and an unavailable judgement keeps every match.
 ## Auto-save Memory Policy
 Configure `auto_save_memory_policy` on `Agent` to decide which events are persisted by automatic long-term-memory saving. If omitted, it is equivalent to `"default"`.
 ```python

diff --git a/docs/content/docs/framework/memory/long-term/index.mdx b/docs/content/docs/framework/memory/long-term/index.mdx
@@ -31,6 +31,8 @@ ltm = LongTermMemory(backend="viking", app_name="ltm_demo")
 | `index` | `str` | `""` | 存储记忆所用的索引/集合名。为空时回退到 `app_name`，再为空则用 `default_app`。 |
 | `app_name` | `str` | `""` | 拥有该记忆的应用名，常用作数据隔离与 `index` 的回退值。 |
 | `user_id` | `str` | `""` | **已废弃**，仅为向后兼容保留。 |
+| `recall_strategy` | `str` | `"off"` | 召回判定策略；`decision` 时由判定模型过滤与本次请求无关的记忆。 |
+| `recall_relevance_threshold` | `float` | `0.5` | 召回相关度阈值；判定低于该值的记忆不返回。 |
 
 <Callout type="info">
 向量类后端（`local`、`opensearch`、`redis`）会对记忆做向量化（embedding），需要安装扩展依赖：`pip install "veadk-python[extensions]"`，并配置 embedding 模型（环境变量前缀 `MODEL_EMBEDDING_`，缺省时复用 `MODEL_AGENT_API_KEY`）。`viking`、`mem0`、`openviking` 与 `tos_context` 是托管服务，无需本地 embedding。
@@ -173,6 +175,17 @@ agent = Agent(
 ```
 
 为避免索引被频繁初始化，VeADK 提供 `MIN_MESSAGES_THRESHOLD` 与 `MIN_TIME_THRESHOLD` 两个环境变量自定义保存周期：默认在累计 10 条 event 或间隔 60 秒时触发保存；此外，当切换 `session_id` 并发起新问答时，VeADK 会自动把上一个会话写入长期记忆。
+
+默认保存策略是 `MEMORY_SAVE_STRATEGY=threshold`，即上面的双阈值规则。设置为 `decision` 后，改由已配置的判定模型判断"这一轮是否包含值得长期记住的内容"（陈述过的偏好、事实、决策、约束等），阈值由 `MEMORY_SAVE_WORTH_THRESHOLD` 控制，默认 `0.5`：
+
+| 判定结果 | 行为 |
+| --- | --- |
+| 通过阈值 | 立即写入，不必等满 10 条 event 或 60 秒 |
+| 未通过阈值 | 跳过，不会仅因为攒够条数就写入 |
+| 判定不可用 | 回落到 `MIN_MESSAGES_THRESHOLD` / `MIN_TIME_THRESHOLD` |
+
+被跳过的轮次不会推进保存游标，后续判定通过时会把这批 event 一起写入，因此不会丢数据。判定模型的环境变量见 [环境变量参考](/references/configuration/environment-variables)。
+检索侧同样可以交给判定模型：`MEMORY_RECALL_STRATEGY=decision` 时，`search_memory` 返回前会逐条给召回片段打相关度（不相关 / 同主题但用不上 / 有用背景 / 必须遵守），低于 `MEMORY_RECALL_RELEVANCE_THRESHOLD`（默认 `0.5`）的会被丢弃；一次判定最多覆盖 20 条，超出的按后端顺序原样保留，判定不可用时保留全部结果。
 ## 自动保存记忆策略
 开发者在 `Agent` 上配置 `auto_save_memory_policy` 来控制自动保存长期记忆时哪些 event 会被写入；不配置时等价于 `"default"`。
 ```python

diff --git a/docs/content/docs/references/configuration/environment-variables.en.mdx b/docs/content/docs/references/configuration/environment-variables.en.mdx
@@ -2,7 +2,7 @@
 title: "Environment variables"
 ---
 
-Environment variables are the primary configuration method, supplying secrets, connection details, and service addresses. Names are grouped by domain prefix: Volcengine account, models, built-in tools, databases, and observability. Every variable can also be expressed as the equivalent nested structure in `config.yaml`.
+Environment variables are the primary configuration method, supplying secrets, connection details, and service addresses. Names are grouped by domain prefix: Volcengine account, models, built-in tools, Harness Extension, decision model, long-term memory, databases, and observability. Every variable can also be expressed as the equivalent nested structure in `config.yaml`.
 
 <Callout type="info">
 Supply secrets through environment variables or `config.yaml`. The tables below list common variables by domain; configure only those for the capabilities actually used.
@@ -76,9 +76,68 @@ Prefix `HARNESS_`, used to attach optional Harness plugins to HarnessApp Runtime
 | `HARNESS_MAX_TOOL_RESULT_CHARS` | Single tool-result compaction threshold, default `4000`. |
 | `HARNESS_VERIFIER_MODE` | Final-response verification mode, `observe` or `block`; default `observe`. |
 | `HARNESS_STORE_PATH` | Optional JSONL event store path; in-memory store is used when unset. |
+| `HARNESS_COMPACTION_STRATEGY` | Compaction candidate strategy, `builtin` or `decision`; default `builtin`. |
+| `HARNESS_LONG_RUN_STRATEGY` | Long-run steering strategy, `counter` or `decision`; default `counter`. |
+| `HARNESS_MODE_STRATEGY` | Context mode-block strategy, `keywords` or `decision`; default `keywords`. |
+| `HARNESS_VERIFIER_STRATEGY` | Final-answer verification, `deterministic` or `decision`; default `deterministic`. |
+| `HARNESS_COMPACTION_KEEP_THRESHOLD` | Compaction candidates: a candidate is kept when the judged probability is at or above this value; default `0.5`. |
+| `HARNESS_LONG_RUN_READY_THRESHOLD` | Long-run steering: guidance is injected when the judged probability of being ready is at or above this value; default `0.5`. |
+| `HARNESS_LONG_RUN_MIN_CONFIDENCE` | Smallest confidence a judged steering action needs; below it the plugin keeps the default wording and only the convergence probability counts; default `0` (off). |
+| `HARNESS_MODE_DECISION_THRESHOLD` | Context mode blocks: a block is injected when the judged probability is at or above this value; default `0.5`. |
+| `HARNESS_VERIFIER_SUPPORT_THRESHOLD` | Final-answer support: the answer fails when the judged support is below this value; default `0.5`. |
+| `HARNESS_VERIFIER_OVERCLAIM_THRESHOLD` | Final-answer support: the answer fails when the judged probability of claiming more than the receipts show is at or above this value, even when the verdict was `supported`; default `0.5`. |
+| `HARNESS_VERIFIER_MIN_CONFIDENCE` | Smallest confidence a judged verdict needs; below it the judge gives none and the builtin rules decide; default `0` (off). |
+| `HARNESS_SKILL_STRATEGY` | Advertised-skills strategy, `all` or `decision`; default `all`. Requires the `skill_prefilter` component. |
+| `HARNESS_SKILL_DECISION_THRESHOLD` | Advertised skills: a skill stays in the request when the judged probability of needing it is at or above this value; default `0.5`. |
+| `HARNESS_SKILL_MAX_CANDIDATES` | Advertised skills: a list longer than this is not judged, so every skill stays advertised; default `40`. |
+| `HARNESS_ROUTING_STRATEGY` | Sub-agent routing strategy, `model` or `decision`; default `model`. Requires the `agent_routing` component. |
+| `HARNESS_ROUTING_DECISION_THRESHOLD` | Routing: the judged agent is transferred to only at or above this probability; default `0.5`. |
 
 The `harness_enhance` block maps to these environment variables when deploying a HarnessApp Runtime. Prefer `harness.yaml` or `veadk agentkit invoke` flags for normal developer workflows; use environment variables for platform integration and container runtimes.
 
+## Decision model
+
+Prefix `DECISION_MODEL_`, for attaching the optional judgement model (Jev). It is fully independent of the agent model and serves a small number of judgement points; when it is unconfigured or a call fails, each caller falls back to its built-in rule. Both `enabled` and `api_key` are required for the model to count as configured.
+
+| Variable | Meaning |
+| :- | :- |
+| `DECISION_MODEL_ENABLED` | Enable the judgement model; accepts `true` / `false`; default `false`. |
+| `DECISION_MODEL_PROVIDER` | Service provider, `typesafe` / `openrouter` / `systemone`; default `typesafe`. An unknown value is treated as `typesafe`. |
+| `DECISION_MODEL_NAME` | Model name, default `jev-latest`. |
+| `DECISION_MODEL_API_BASE` | Service address; defaults per provider when unset, and `/v1/systemone` is appended when missing. |
+| `DECISION_MODEL_API_KEY` | Access key; an empty value counts as unconfigured. |
+| `DECISION_MODEL_TIMEOUT` | Budget for one judgement in seconds, retries and backoff included; default `5`, ceiling `5`. |
+| `DECISION_MODEL_MAX_RETRIES` | Retries per judgement request, default `3`. |
+| `DECISION_MODEL_FAILURE_THRESHOLD` | Consecutive failures that trip the circuit breaker, default `3`; `0` disables it. |
+| `DECISION_MODEL_COOLDOWN_SECONDS` | Cooldown length in seconds after tripping, default `30`. |
+
+```yaml title="config.yaml"
+model:
+  decision:
+    enabled: true
+    provider: openrouter
+    name: typesafe/jev-1.13
+    # api_key:
+    # timeout: 5
+```
+
+`model.decision.*` in `config.yaml` expands to `MODEL_DECISION_*`, which is equivalent to `DECISION_MODEL_*`; when both are set, `DECISION_MODEL_*` wins.
+
+When deploying a HarnessApp Runtime, the top-level `decision_model` block in `harness.yaml` maps to this group.
+
+## Long-term memory
+
+Long-term memory has one optional judgement point for saving and one for recall. Both strategies are read at module import time, so a change requires a process restart; an unconfigured or unavailable judgement falls back to the previous behaviour.
+
+| Kind | Variable | Meaning |
+| :- | :- | :- |
+| Save | `MEMORY_SAVE_STRATEGY` | `threshold` uses the numeric thresholds only; `decision` lets the judgement model decide; default `threshold`. |
+| | `MEMORY_SAVE_WORTH_THRESHOLD` | Probability above which a turn is worth remembering, default `0.5`; below it the save is skipped and the cursor does not advance. |
+| | `MIN_MESSAGES_THRESHOLD` | Minimum new messages under the `threshold` strategy, default `10`. |
+| | `MIN_TIME_THRESHOLD` | Minimum seconds between saves under the `threshold` strategy, default `60`. |
+| Recall | `MEMORY_RECALL_STRATEGY` | `off` returns the backend ranking as-is; `decision` rates each memory and drops the low ones; default `off`. |
+| | `MEMORY_RECALL_RELEVANCE_THRESHOLD` | Relevance threshold, default `0.5`; memories below it are dropped. Memories the judgement did not rate, and those beyond the 20-item limit, are kept as-is. |
+
 ## Databases
 
 Prefix `DATABASE_`, grouped by storage type. Memory and the knowledge base read the matching config based on the selected backend.

diff --git a/docs/content/docs/references/configuration/environment-variables.mdx b/docs/content/docs/references/configuration/environment-variables.mdx
@@ -2,7 +2,7 @@
 title: "环境变量"
 ---
 
-环境变量是最主要的配置方式，提供密钥、连接信息与服务地址。变量名按领域前缀分组：火山引擎账号、模型、内置工具、数据库、可观测各成一类。每个环境变量都能在 `config.yaml` 中以对应层级结构表达，两者等价。
+环境变量是最主要的配置方式，提供密钥、连接信息与服务地址。变量名按领域前缀分组：火山引擎账号、模型、内置工具、Harness Extension、决策模型、长记忆、数据库、可观测各成一类。每个环境变量都能在 `config.yaml` 中以对应层级结构表达，两者等价。
 
 <Callout type="info">
 密钥统一通过环境变量或 `config.yaml` 提供。下表按领域列出常用变量，仅需配置实际用到的能力对应项。
@@ -76,9 +76,68 @@ volcengine:
 | `HARNESS_MAX_TOOL_RESULT_CHARS` | 单个工具结果压缩阈值，默认 `4000`。 |
 | `HARNESS_VERIFIER_MODE` | 最终回答校验模式，`observe` 或 `block`，默认 `observe`。 |
 | `HARNESS_STORE_PATH` | 可选 JSONL 事件存储路径；不设置时使用内存存储。 |
+| `HARNESS_COMPACTION_STRATEGY` | 工具结果压缩候选策略，`builtin` 或 `decision`，默认 `builtin`。 |
+| `HARNESS_LONG_RUN_STRATEGY` | 长任务收尾引导策略，`counter` 或 `decision`，默认 `counter`。 |
+| `HARNESS_MODE_STRATEGY` | 上下文模式块策略，`keywords` 或 `decision`，默认 `keywords`。 |
+| `HARNESS_VERIFIER_STRATEGY` | 最终回答校验策略，`deterministic` 或 `decision`，默认 `deterministic`。 |
+| `HARNESS_COMPACTION_KEEP_THRESHOLD` | 压缩候选保留阈值，默认 `0.5`；判定保留概率低于该值即压缩。 |
+| `HARNESS_LONG_RUN_READY_THRESHOLD` | 长任务收尾阈值，默认 `0.5`；判定可收尾概率高于该值即注入引导。 |
+| `HARNESS_LONG_RUN_MIN_CONFIDENCE` | 长任务引导动作的最低置信度，默认 `0`（关闭）；判定对该动作没把握时只用默认措辞，保留「还没收敛」的信号。 |
+| `HARNESS_MODE_DECISION_THRESHOLD` | 上下文模式块阈值，默认 `0.5`；判定概率高于该值即注入对应模式块。 |
+| `HARNESS_VERIFIER_SUPPORT_THRESHOLD` | 最终回答支撑度阈值，默认 `0.5`；判定支撑度低于该值即判为失败。 |
+| `HARNESS_VERIFIER_OVERCLAIM_THRESHOLD` | 回答超出回执范围的否决阈值，默认 `0.5`；判定概率不低于该值直接判为失败，即使结论是 `supported`。 |
+| `HARNESS_VERIFIER_MIN_CONFIDENCE` | 最终回答判定的最低置信度，默认 `0`（关闭）；低于该值时不做判定，回落到内置规则。 |
+| `HARNESS_SKILL_STRATEGY` | 技能广告策略，`all` 或 `decision`，默认 `all`；需要 `skill_prefilter` 组件。 |
+| `HARNESS_SKILL_DECISION_THRESHOLD` | 技能广告阈值，默认 `0.5`；判定需要该技能的概率不低于该值才继续广告。 |
+| `HARNESS_SKILL_MAX_CANDIDATES` | 技能候选上限，默认 `40`；技能数量超过该值时不判定，全部照常广告。 |
+| `HARNESS_ROUTING_STRATEGY` | 子 Agent 路由策略，`model` 或 `decision`，默认 `model`；需要 `agent_routing` 组件。 |
+| `HARNESS_ROUTING_DECISION_THRESHOLD` | 路由阈值，默认 `0.5`；判定概率不低于该值才直接转移。 |
 
 `harness_enhance` 配置块会在 HarnessApp Runtime 部署时映射为这些环境变量。推荐开发者优先通过 `harness.yaml` 或 `veadk agentkit invoke` 参数启用，环境变量适合平台集成和镜像运行时。
 
+## 决策模型
+
+统一前缀 `DECISION_MODEL_`，用于挂载可选的判定模型（Jev）。判定模型与对话模型完全独立，只服务少数判定点；未配置或调用失败时，各调用方回落到原内置规则。`enabled` 与 `api_key` 同时具备才算配置完成。
+
+| 环境变量 | 释义 |
+| :- | :- |
+| `DECISION_MODEL_ENABLED` | 是否启用判定模型，支持 `true` / `false`，默认 `false`。 |
+| `DECISION_MODEL_PROVIDER` | 服务提供方，`typesafe` / `openrouter` / `systemone`，默认 `typesafe`；未知取值按 `typesafe` 处理。 |
+| `DECISION_MODEL_NAME` | 模型名称，默认 `jev-latest`。 |
+| `DECISION_MODEL_API_BASE` | 服务地址；不设置时按 provider 取默认值，未以 `/v1/systemone` 结尾会自动补上。 |
+| `DECISION_MODEL_API_KEY` | 访问密钥；为空视为未配置。 |
+| `DECISION_MODEL_TIMEOUT` | 单次判定预算（秒，含重试与退避），默认 `5`，上限 `5`。 |
+| `DECISION_MODEL_MAX_RETRIES` | 判定请求重试次数，默认 `3`。 |
+| `DECISION_MODEL_FAILURE_THRESHOLD` | 连续失败多少次后熔断，默认 `3`；`0` 表示不熔断。 |
+| `DECISION_MODEL_COOLDOWN_SECONDS` | 熔断后的冷却时长（秒），默认 `30`。 |
+
+```yaml title="config.yaml"
+model:
+  decision:
+    enabled: true
+    provider: openrouter
+    name: typesafe/jev-1.13
+    # api_key:
+    # timeout: 5
+```
+
+`config.yaml` 的 `model.decision.*` 会被展开为 `MODEL_DECISION_*`，与 `DECISION_MODEL_*` 等价；两者同时设置时 `DECISION_MODEL_*` 优先。
+
+HarnessApp Runtime 部署时，`harness.yaml` 的顶层 `decision_model` 块映射为这组变量。
+
+## 长记忆
+
+长记忆的落库与召回各有一个可选判定点。策略都在模块导入时读取，修改后需重启进程；判定不可用或未配置时，回落原有行为。
+
+| 子类 | 环境变量 | 释义 |
+| :- | :- | :- |
+| 落库判定 | `MEMORY_SAVE_STRATEGY` | `threshold` 只走数值阈值；`decision` 交判定模型决定，默认 `threshold`。 |
+| | `MEMORY_SAVE_WORTH_THRESHOLD` | 值得长期记住的概率阈值，默认 `0.5`；低于该值跳过写入、且不推进游标。 |
+| | `MIN_MESSAGES_THRESHOLD` | `threshold` 策略下最少新增消息数，默认 `10`。 |
+| | `MIN_TIME_THRESHOLD` | `threshold` 策略下两次写入的最小间隔（秒），默认 `60`。 |
+| 召回判定 | `MEMORY_RECALL_STRATEGY` | `off` 直接返回后端排序结果；`decision` 逐条评估相关度并丢弃低分项，默认 `off`。 |
+| | `MEMORY_RECALL_RELEVANCE_THRESHOLD` | 相关度阈值，默认 `0.5`；低于该值丢弃。未评估到的记忆、以及超出 20 条判定上限的部分原样保留。 |
+
 ## 数据库
 
 统一前缀 `DATABASE_`，按存储类型分组。记忆与知识库按所选后端读取对应配置。