feat(memory): relevance-gated auto recall with hard timeout

Replace the unconditional top-5 keyword recall with a deterministic,
layered gate so only relevant Knowledge entries reach the prompt.

- New src/memory/recall.rs: tokenization with stopword filtering and a
  bounded term list; ranking combines lexical relevance, importance and
  recency; double gate (min_relevance + min_score) drops weak or stale
  matches; hard tokio::time::timeout wraps the SQL search so a slow FTS5
  query never delays a turn.
- MemoryConfig gains recall_min_relevance, recall_min_score,
  recall_recency_half_life_days, recall_timeout_ms (default 1000ms);
  recall_limit is now actually wired instead of being a dead field.
- MemoryManager::recall_for_context exposes the gated path; the
  memory_recall tool keeps the raw search.
- Storage::search_memories / search_memories_by_time share a single
  jieba-based tokenizer via the new module; search_memories_by_terms
  accepts pre-tokenized input for the gated path.
- Docs and example configs (README, about-picobot references, both
  config.example.json templates) updated to reflect the new behavior.

README.md

@@ -329,7 +329,7 @@ PicoBot 有两类记忆：
 | Knowledge | 偏好、事实、项目规则、长期可复用信息 | 长期保留，手动删除 |
 | Timeline | 长对话压缩后的历史摘要 | 默认保留 90 天 |
 
-每轮处理用户消息时，MemoryManager 会按用户输入召回 Knowledge，并作为运行时上下文附加到本轮用户消息。当前召回上限固定为 5；`memory.recall_limit` 已支持解析但尚未接入 worker。长会话使用一个活动 checkpoint：累计摘要加 `first_retained_seq` 之后的原始消息尾部构成模型上下文，原始消息、工具调用结果、ID 和 seq 均不会被压缩改写。旧工具结果会保留在原始历史中，但 checkpoint 边界推进后不再永久占用 Provider 上下文。成功的语义摘要还会 best-effort 保存为 Timeline，供 `timeline_recall` 检索；Timeline 不参与会话恢复正确性。Scheduler 默认创建一个每日维护任务，按 `memory.timeline_retention_days` 清理过期 Timeline；结果通过 `complete_scheduled_run` 结构化提交，Knowledge 不会被自动删除。
+每轮处理用户消息时，MemoryManager 会按用户输入召回 Knowledge，并作为运行时上下文附加到本轮用户消息。自动召回是确定性的关键词检索（jieba 分词 + FTS5），按「词项相关度 + 重要度 + 时效」加权并通过相关性/综合分双门槛过滤，条数受 `memory.recall_limit` 约束，搜索受 `memory.recall_timeout_ms` 硬超时保护，超时或无关时本轮不注入。长会话使用一个活动 checkpoint：累计摘要加 `first_retained_seq` 之后的原始消息尾部构成模型上下文，原始消息、工具调用结果、ID 和 seq 均不会被压缩改写。旧工具结果会保留在原始历史中，但 checkpoint 边界推进后不再永久占用 Provider 上下文。成功的语义摘要还会 best-effort 保存为 Timeline，供 `timeline_recall` 检索；Timeline 不参与会话恢复正确性。Scheduler 默认创建一个每日维护任务，按 `memory.timeline_retention_days` 清理过期 Timeline；结果通过 `complete_scheduled_run` 结构化提交，Knowledge 不会被自动删除。
 
 模型的 `models.<name>.token_limit` 给出上下文窗口上限，未配置时默认为 128,000；Agent 的 `agents.<name>.token_limit` 是可选的收紧上限，两者都有配置时有效窗口取二者最小值，因此 Agent 不能扩大模型窗口。自动压缩使用保留量阈值 `context_tokens > context_window - effective_reserve`，默认 reserve 为 16,384 tokens，并尽量原样保留最近 20,000 tokens。小窗口会自动把 reserve 限制为窗口的一半、把近期保留量限制为有效阈值的一半。摘要请求不使用固定 32K 输入上限，而是按有效窗口扣除摘要输出、提示词和安全余量；超大历史只在摘要请求副本中按“已有 checkpoint + 最新消息优先”生成有界 head/tail 转录，SQLite 原文不变。手动 `/compact` 跳过自动阈值；换成小模型后若发送前预检已发现硬超限，或首次请求返回真实 context overflow，语义摘要不可用时才使用明确标记的确定性降级裁剪，正式请求最多重试一次。若 overflow 发生在工具已经执行之后，AgentLoop 只在当前内存转录上裁掉旧完整 Turn 并重试当前模型步骤一次，不会从数据库历史重跑工具。
 
@@ -403,7 +403,11 @@ Skill 是包含 `SKILL.md` 的目录。加载优先级从高到低：
 | `context_compaction.enabled` | `true` |
 | `context_compaction.reserve_tokens` | `16384` |
 | `context_compaction.keep_recent_tokens` | `20000` |
-| `memory.recall_limit` | `5`（当前运行时固定为 5） |
+| `memory.recall_limit` | `5` |
+| `memory.recall_min_relevance` | `0.25` |
+| `memory.recall_min_score` | `0.25` |
+| `memory.recall_recency_half_life_days` | `30` |
+| `memory.recall_timeout_ms` | `1000` |
 | `memory.timeline_retention_days` | `90` |
 | `mcp.tool_timeout_secs` | `180` |
 | `mcp.servers[].tool_settings` | `{}`；可按工具名声明 `read_only` / `exclusive`，并发状态自动推导 |

resources/skills/about-picobot/assets/config.example.json

@@ -87,6 +87,10 @@
         "consolidation_provider": null,
         "consolidation_model": null,
         "recall_limit": 5,
+        "recall_min_relevance": 0.25,
+        "recall_min_score": 0.25,
+        "recall_recency_half_life_days": 30,
+        "recall_timeout_ms": 1000,
         "idle_consolidation_minutes": 10,
         "timeline_retention_days": 90,
         "max_failures_before_degrade": 3

resources/skills/about-picobot/references/architecture.md

@@ -228,7 +228,7 @@ WebUI/TUI 的 Active Turn 使用 `send_message(files=...)` 向自身 session 投
 | 语义 checkpoint 提交后 | 摘要 best-effort 存储为 Timeline 记忆 |
 | 会话恢复 | 从 checkpoint 与原始 seq 确定性重建，不读取 Timeline |
 
-`memory.recall_limit`、`idle_consolidation_minutes`、`timeline_retention_days` 和 `max_failures_before_degrade` 当前会被配置解析；其中每轮 Knowledge 召回在 worker 中仍固定为 5，其余自动维护策略尚未接入运行循环。不要把“配置可解析”误认为“行为已生效”。
+`memory.recall_limit` 及新增的 `recall_min_relevance`、`recall_min_score`、`recall_recency_half_life_days`、`recall_timeout_ms` 已生效：每轮自动召回用 jieba 分词 + FTS5 检索候选，按「词项相关度 0.5 + 重要度 0.3 + 时效 0.2」加权并通过相关性/综合分双门槛过滤，搜索受硬超时保护。`idle_consolidation_minutes`、`timeline_retention_days` 和 `max_failures_before_degrade` 仍只是配置解析：自动 idle consolidation、Timeline 清理和失败降级循环尚未接入运行循环。不要把“配置可解析”误认为“行为已生效”。
 
 ---
 

resources/skills/about-picobot/references/config.md

@@ -111,12 +111,16 @@ Gateway WebUI 的“配置”页可以编辑实际加载的配置文件。读取
 |------|------|------|------|
 | `consolidation_provider` | string | 主 Agent provider | 当前记录在 MemoryManager 中供后续归并使用；压缩摘要仍使用 Session provider |
 | `consolidation_model` | string | 主 Agent model | 当前记录在 MemoryManager 中供后续归并使用；压缩摘要仍使用 Session model |
-| `recall_limit` | int | 5 | 预期的每轮知识召回上限；当前 worker 固定使用 5 |
+| `recall_limit` | int | 5 | 每轮自动注入上下文的知识记忆条数上限 |
+| `recall_min_relevance` | float | 0.25 | 自动召回的相关性硬门槛（命中词项占比低于此值则丢弃） |
+| `recall_min_score` | float | 0.25 | 自动召回的综合分门槛（相关度+重要度+时效加权） |
+| `recall_recency_half_life_days` | int | 30 | 自动召回时效衰减的半衰期（天） |
+| `recall_timeout_ms` | int | 1000 | 自动召回搜索的硬超时；超时本轮不注入记忆 |
 | `idle_consolidation_minutes` | int | 10 | 预留的空闲归并阈值；当前无对应循环 |
 | `timeline_retention_days` | int | 90 | 默认日常维护巡检删除超过该期限的 Timeline；Knowledge 不受影响 |
 | `max_failures_before_degrade` | int | 3 | 预留的归并失败阈值；当前无失败降级循环 |
 
-注意：当前 worker 的 Knowledge 召回数量仍固定为 5；idle consolidation 和失败降级循环尚未接入。Timeline 清理由默认启用的 `picobot-routine-maintenance` Scheduled Run 执行；该任务使用 `never` 策略，结构化结果只进入运行审计和 Health。
+自动召回每轮用当前用户输入做关键词检索（jieba 分词 + FTS5），按「词项相关度 0.5 + 重要度 0.3 + 时效 0.2」加权，通过相关性/综合分双门槛后才注入；搜索有硬超时保证不拖慢 Turn。Timeline 不自动召回，需显式 `timeline_recall`。idle consolidation 和失败降级循环尚未接入。Timeline 清理由默认启用的 `picobot-routine-maintenance` Scheduled Run 执行；该任务使用 `never` 策略，结构化结果只进入运行审计和 Health。
 
 ## channels.feishu 字段
 

resources/skills/about-picobot/references/faq.md

@@ -64,7 +64,7 @@ LLM 调用记录存储在 `llm_calls` 表中。可通过 SQLite 客户端直接
 
 ## Q: 为什么修改了某些 memory 配置却没有看到行为变化？
 
-当前 `recall_limit`、`idle_consolidation_minutes`、`timeline_retention_days` 和 `max_failures_before_degrade` 都能被配置解析，但每轮 Knowledge 召回仍固定为 5，自动 idle consolidation、Timeline 清理和失败降级循环尚未接入。以当前代码行为为准。
+当前 `recall_limit` 以及新增的 `recall_min_relevance`、`recall_min_score`、`recall_recency_half_life_days`、`recall_timeout_ms` 均已生效，控制每轮自动知识召回的相关性门槛、加权与超时。`idle_consolidation_minutes`、`timeline_retention_days` 和 `max_failures_before_degrade` 仍只是配置解析：自动 idle consolidation、Timeline 清理（由维护任务执行）和失败降级循环尚未接入。以当前代码行为为准。
 
 ## Q: Gateway 为什么无法立即退出？
 

resources/templates/config.example.json

@@ -111,6 +111,10 @@
         "consolidation_provider": null,
         "consolidation_model": null,
         "recall_limit": 5,
+        "recall_min_relevance": 0.25,
+        "recall_min_score": 0.25,
+        "recall_recency_half_life_days": 30,
+        "recall_timeout_ms": 1000,
         "idle_consolidation_minutes": 10,
         "timeline_retention_days": 90,
         "max_failures_before_degrade": 3

src/config/mod.rs

@@ -555,6 +555,21 @@ pub struct MemoryConfig {
     /// Max knowledge entries injected into system prompt per turn.
     #[serde(default = "default_recall_limit")]
     pub recall_limit: usize,
+    /// Minimum lexical relevance (fraction of matched query terms) for a
+    /// memory to be injected during automatic per-turn recall.
+    #[serde(default = "default_recall_min_relevance")]
+    pub recall_min_relevance: f64,
+    /// Minimum combined score (relevance + importance + recency) for a memory
+    /// to be injected during automatic per-turn recall.
+    #[serde(default = "default_recall_min_score")]
+    pub recall_min_score: f64,
+    /// Half-life (days) of the recency decay factor used in automatic recall.
+    #[serde(default = "default_recall_recency_half_life_days")]
+    pub recall_recency_half_life_days: u64,
+    /// Hard timeout (ms) for the automatic recall search; on expiry the turn
+    /// proceeds without injected memories.
+    #[serde(default = "default_recall_timeout_ms")]
+    pub recall_timeout_ms: u64,
     /// Idle minutes before triggering consolidation (for async channels).
     #[serde(default = "default_idle_consolidation_minutes")]
     pub idle_consolidation_minutes: u64,
@@ -572,6 +587,10 @@ impl Default for MemoryConfig {
             consolidation_provider: None,
             consolidation_model: None,
             recall_limit: 5,
+            recall_min_relevance: 0.25,
+            recall_min_score: 0.25,
+            recall_recency_half_life_days: 30,
+            recall_timeout_ms: 1000,
             idle_consolidation_minutes: 10,
             timeline_retention_days: 90,
             max_failures_before_degrade: 3,
@@ -789,6 +808,18 @@ impl Default for BrowserConfig {
 fn default_recall_limit() -> usize {
     5
 }
+fn default_recall_min_relevance() -> f64 {
+    0.25
+}
+fn default_recall_min_score() -> f64 {
+    0.25
+}
+fn default_recall_recency_half_life_days() -> u64 {
+    30
+}
+fn default_recall_timeout_ms() -> u64 {
+    1000
+}
 fn default_idle_consolidation_minutes() -> u64 {
     10
 }

src/gateway/mod.rs

@@ -156,11 +156,16 @@ impl GatewayState {
         let consolidation_model = config
             .memory
             .resolve_consolidation_model(&provider_config.model_id);
-        let memory_manager = Arc::new(MemoryManager::new(
+        let memory_manager = Arc::new(
+            MemoryManager::new(
                 storage.clone(),
                 consolidation_provider,
                 consolidation_model,
-        ));
+            )
+            .with_recall(crate::memory::recall::RecallConfig::from_memory_config(
+                &config.memory,
+            )),
+        );
         tracing::info!(
             consolidation_provider = %memory_manager.consolidation_provider,
             consolidation_model = %memory_manager.consolidation_model,

src/memory/mod.rs

@@ -1,3 +1,4 @@
+pub mod recall;
 pub mod types;
 
 use std::sync::Arc;
@@ -6,6 +7,8 @@ use uuid::Uuid;
 use crate::storage::Storage;
 pub use types::{ConsolidationFact, ConsolidationResult, MemoryCategory, MemoryEntry};
 
+use recall::RecallConfig;
+
 /// MemoryManager provides high-level memory operations.
 /// Wraps the Storage SQLite layer with semantic methods.
 #[derive(Clone)]
@@ -13,6 +16,7 @@ pub struct MemoryManager {
     storage: Arc<Storage>,
     pub consolidation_provider: String,
     pub consolidation_model: String,
+    recall: RecallConfig,
 }
 
 impl MemoryManager {
@@ -25,9 +29,57 @@ impl MemoryManager {
             storage,
             consolidation_provider,
             consolidation_model,
+            recall: RecallConfig::default(),
         }
     }
 
+    /// Override the automatic-recall knobs (derived from `config.memory`).
+    pub fn with_recall(mut self, config: RecallConfig) -> Self {
+        self.recall = config;
+        self
+    }
+
+    /// Automatic per-turn recall for context injection.
+    ///
+    /// Returns up to `recall.limit` Knowledge entries that clear both the
+    /// relevance and combined-score gates. The search is bounded by a hard
+    /// timeout: on expiry, error, or an empty normalized query it returns an
+    /// empty vector so the turn is never delayed by memory lookup.
+    pub async fn recall_for_context(&self, query: &str) -> Vec<MemoryEntry> {
+        let terms = recall::normalize_query(query);
+        if terms.is_empty() {
+            return Vec::new();
+        }
+        let candidate_limit = (self.recall.limit * recall::CANDIDATE_FACTOR)
+            .min(recall::MAX_CANDIDATES)
+            .max(self.recall.limit);
+        let candidates = tokio::time::timeout(
+            self.recall.timeout,
+            self.storage.search_memories_by_terms(
+                &terms,
+                Some(&MemoryCategory::Knowledge),
+                None,
+                candidate_limit,
+            ),
+        )
+        .await;
+        let candidates = match candidates {
+            Ok(Ok(entries)) => entries,
+            Ok(Err(error)) => {
+                tracing::debug!(error = %error, "memory recall failed; skipping injection");
+                return Vec::new();
+            }
+            Err(_) => {
+                tracing::debug!("memory recall timed out; skipping injection");
+                return Vec::new();
+            }
+        };
+        let now_ms = chrono::Utc::now().timestamp_millis();
+        let mut ranked = recall::rank_and_gate(candidates, &terms, &self.recall, now_ms);
+        ranked.truncate(self.recall.limit);
+        ranked
+    }
+
     /// Store or update a memory entry. Generates timestamp and UUID.
     pub async fn store(
         &self,
@@ -88,7 +140,10 @@ impl MemoryManager {
 
     /// Check if the memory system has any entries (for testing/health check).
     pub async fn is_empty(&self) -> Result<bool, crate::storage::StorageError> {
-        self.recall("*", 1, None, None).await.map(|r| r.is_empty())
+        self.storage
+            .list_memories(None, None, 1)
+            .await
+            .map(|entries| entries.is_empty())
     }
 }
 
@@ -257,4 +312,65 @@ mod tests {
         assert_eq!(scoped[0].key, "tl_a");
         assert_eq!(scoped[0].session_id.as_deref(), Some("chan:chat:dialog_a"));
     }
+
+    #[tokio::test]
+    async fn recall_for_context_returns_empty_for_stopword_only_query() {
+        let (mm, _dir) = setup_memory_manager().await;
+        mm.store(
+            "user_pref",
+            "user prefers python",
+            MemoryCategory::Knowledge,
+            None,
+            Some(0.9),
+        )
+        .await
+        .unwrap();
+        // A query made entirely of stopwords must not surface unrelated memories.
+        assert!(mm.recall_for_context("你好吗请吧").await.is_empty());
+    }
+
+    #[tokio::test]
+    async fn recall_for_context_recalls_relevant_and_gates_irrelevant() {
+        let (mm, _dir) = setup_memory_manager().await;
+        mm.store(
+            "user_pref",
+            "user prefers python for scripting",
+            MemoryCategory::Knowledge,
+            None,
+            Some(0.9),
+        )
+        .await
+        .unwrap();
+        mm.store(
+            "favorite_color",
+            "user favorite color is blue",
+            MemoryCategory::Knowledge,
+            None,
+            Some(0.9),
+        )
+        .await
+        .unwrap();
+
+        let hits = mm.recall_for_context("帮我写一个 python 脚本").await;
+        assert_eq!(hits.len(), 1);
+        assert_eq!(hits[0].key, "user_pref");
+    }
+
+    #[tokio::test]
+    async fn recall_for_context_respects_limit() {
+        let (mm, _dir) = setup_memory_manager().await;
+        for i in 0..8 {
+            mm.store(
+                &format!("pref_{i}"),
+                format!("user preference about python number {i}").as_str(),
+                MemoryCategory::Knowledge,
+                None,
+                Some(0.9),
+            )
+            .await
+            .unwrap();
+        }
+        let hits = mm.recall_for_context("python preference").await;
+        assert!(hits.len() <= 5);
+    }
 }

src/memory/recall.rs

@@ -0,0 +1,242 @@
+//! Deterministic, relevance-gated memory recall for per-turn context injection.
+//!
+//! This is the *automatic* recall path (as opposed to the user-facing
+//! `memory_recall` tool, which performs an unfiltered search). It tokenizes
+//! and normalizes the query, retrieves a small candidate set through FTS5,
+//! then ranks and gates candidates by lexical relevance, importance, and
+//! recency. Everything here is bounded and synchronous so a turn is never
+//! delayed by a slow search.
+
+use std::collections::HashSet;
+use std::sync::OnceLock;
+use std::time::Duration;
+
+use jieba_rs::Jieba;
+
+use super::MemoryEntry;
+
+/// Weights for the combined score (sum to 1.0 for a clean 0..1 range).
+const WEIGHT_RELEVANCE: f64 = 0.5;
+const WEIGHT_IMPORTANCE: f64 = 0.3;
+const WEIGHT_RECENCY: f64 = 0.2;
+
+/// Candidate set is `limit * CANDIDATE_FACTOR`, capped by `MAX_CANDIDATES`.
+pub(crate) const CANDIDATE_FACTOR: usize = 4;
+pub(crate) const MAX_CANDIDATES: usize = 50;
+
+/// Hard bounds on the query to keep tokenization and the FTS5 MATCH cheap.
+const MAX_QUERY_CHARS: usize = 512;
+const MAX_TERMS: usize = 12;
+
+const STOPWORDS: &[&str] = &[
+    "的", "了", "是", "在", "我", "你", "他", "她", "它", "们", "吗", "呢", "啊", "吧", "这", "那",
+    "什么", "怎么", "为什么", "帮我", "一下", "请", "要", "想", "能", "可以", "这个", "那个", "今天",
+    "现在", "以及", "还是", "因为", "所以", "如果", "但是", "就", "都", "也", "还", "很", "被", "把",
+    "和", "与", "或", "有", "没有", "一个", "一种", "一些", "哪些", "如何",
+    "a", "an", "the", "to", "of", "for", "and", "or", "what", "how", "please", "can", "do", "is",
+    "are", "my", "you", "i", "we", "this", "that", "me", "it", "on", "in", "at", "with", "about",
+];
+
+/// Runtime knobs for automatic recall, derived from `config.memory`.
+#[derive(Debug, Clone)]
+pub struct RecallConfig {
+    pub limit: usize,
+    pub min_relevance: f64,
+    pub min_score: f64,
+    pub recency_half_life_days: f64,
+    pub timeout: Duration,
+}
+
+impl Default for RecallConfig {
+    fn default() -> Self {
+        Self {
+            limit: 5,
+            min_relevance: 0.25,
+            min_score: 0.25,
+            recency_half_life_days: 30.0,
+            timeout: Duration::from_millis(1000),
+        }
+    }
+}
+
+impl RecallConfig {
+    pub fn from_memory_config(memory: &crate::config::MemoryConfig) -> Self {
+        Self {
+            limit: memory.recall_limit,
+            min_relevance: memory.recall_min_relevance,
+            min_score: memory.recall_min_score,
+            recency_half_life_days: memory.recall_recency_half_life_days as f64,
+            timeout: Duration::from_millis(memory.recall_timeout_ms),
+        }
+    }
+}
+
+fn jieba() -> &'static Jieba {
+    static INSTANCE: OnceLock<Jieba> = OnceLock::new();
+    INSTANCE.get_or_init(Jieba::new)
+}
+
+/// Tokenize with jieba and drop tokens that are too short to be meaningful.
+pub(crate) fn tokenize(query: &str) -> Vec<String> {
+    jieba()
+        .cut(query, true)
+        .into_iter()
+        .map(|token| token.word)
+        .filter(|word| word.len() > 1 || word.bytes().any(|b| b > 127))
+        .map(str::to_string)
+        .collect()
+}
+
+/// Normalize a user query into deduplicated, stopword-free search terms.
+pub(crate) fn normalize_query(query: &str) -> Vec<String> {
+    let bounded: String = if query.chars().count() > MAX_QUERY_CHARS {
+        query.chars().take(MAX_QUERY_CHARS).collect()
+    } else {
+        query.to_string()
+    };
+    let mut seen = HashSet::new();
+    let mut terms: Vec<String> = Vec::new();
+    for term in tokenize(&bounded) {
+        let lower = term.to_lowercase();
+        if STOPWORDS.contains(&lower.as_str()) {
+            continue;
+        }
+        if seen.insert(lower) {
+            terms.push(term);
+        }
+    }
+    terms.truncate(MAX_TERMS);
+    terms
+}
+
+/// Fraction of distinct query terms present in the entry's key or content.
+fn relevance(entry: &MemoryEntry, terms: &[String]) -> f64 {
+    let key = entry.key.to_lowercase();
+    let content = entry.content.to_lowercase();
+    let matched = terms
+        .iter()
+        .filter(|term| {
+            let term = term.to_lowercase();
+            key.contains(&term) || content.contains(&term)
+        })
+        .count();
+    matched as f64 / terms.len() as f64
+}
+
+/// Exponential recency decay in [0,1]; 1.0 for just-updated, 0.0 for very old.
+fn recency(updated_at: &str, half_life_days: f64, now_ms: i64) -> f64 {
+    let Ok(dt) = chrono::DateTime::parse_from_rfc3339(updated_at) else {
+        return 0.0;
+    };
+    let half_life = half_life_days.max(1.0);
+    let age_ms = (now_ms - dt.timestamp_millis()).max(0);
+    let age_days = age_ms as f64 / 86_400_000.0;
+    (-age_days / half_life).exp()
+}
+
+/// Rank candidates by combined score and drop those below the gates.
+pub(crate) fn rank_and_gate(
+    candidates: Vec<MemoryEntry>,
+    terms: &[String],
+    config: &RecallConfig,
+    now_ms: i64,
+) -> Vec<MemoryEntry> {
+    let mut scored: Vec<(f64, f64, MemoryEntry)> = Vec::with_capacity(candidates.len());
+    for entry in candidates {
+        let rel = relevance(&entry, terms);
+        if rel < config.min_relevance {
+            continue;
+        }
+        let importance = entry.importance.clamp(0.0, 1.0);
+        let rec = recency(&entry.updated_at, config.recency_half_life_days, now_ms);
+        let score = WEIGHT_RELEVANCE * rel + WEIGHT_IMPORTANCE * importance + WEIGHT_RECENCY * rec;
+        if score < config.min_score {
+            continue;
+        }
+        scored.push((score, importance, entry));
+    }
+    scored.sort_by(|a, b| {
+        b.0.partial_cmp(&a.0)
+            .unwrap_or(std::cmp::Ordering::Equal)
+            .then_with(|| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal))
+    });
+    scored.into_iter().map(|(_, _, entry)| entry).collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::memory::MemoryCategory;
+
+    fn entry(key: &str, content: &str, importance: f64, updated_at: &str) -> MemoryEntry {
+        MemoryEntry {
+            id: format!("id-{key}"),
+            key: key.to_string(),
+            content: content.to_string(),
+            category: MemoryCategory::Knowledge,
+            importance,
+            session_id: None,
+            created_at: updated_at.to_string(),
+            updated_at: updated_at.to_string(),
+        }
+    }
+
+    #[test]
+    fn normalize_query_removes_stopwords_and_deduplicates() {
+        let terms = normalize_query("请帮我写一个 Python 脚本，用 python 处理");
+        assert!(!terms.contains(&"请".to_string()));
+        assert!(!terms.contains(&"一个".to_string()));
+        // "python" appears twice but is deduplicated (case-insensitive).
+        assert_eq!(terms.iter().filter(|t| t.to_lowercase() == "python").count(), 1);
+    }
+
+    #[test]
+    fn normalize_query_returns_empty_for_all_stopwords() {
+        assert!(normalize_query("呢吗啊吧").is_empty());
+    }
+
+    #[test]
+    fn relevance_counts_distinct_term_matches() {
+        let e = entry("py", "the user prefers python", 0.5, "2024-01-01T00:00:00Z");
+        let terms = vec!["python".to_string(), "rust".to_string()];
+        assert!((relevance(&e, &terms) - 0.5).abs() < 1e-9);
+    }
+
+    #[test]
+    fn gate_drops_low_relevance() {
+        let e = entry("unrelated", "something else entirely", 1.0, "2024-01-01T00:00:00Z");
+        let terms = vec!["python".to_string(), "script".to_string()];
+        let cfg = RecallConfig::default();
+        let out = rank_and_gate(vec![e], &terms, &cfg, 1_704_067_200_000);
+        assert!(out.is_empty());
+    }
+
+    #[test]
+    fn gate_keeps_high_relevance_and_sorts_by_score() {
+        let strong = entry("strong", "python script", 0.5, "2024-01-01T00:00:00Z");
+        let weak = entry("weak", "a python note", 0.9, "2024-01-01T00:00:00Z");
+        let terms = vec!["python".to_string(), "script".to_string()];
+        let cfg = RecallConfig::default();
+        let out = rank_and_gate(vec![weak, strong], &terms, &cfg, 1_704_067_200_000);
+        assert_eq!(out.len(), 2);
+        assert_eq!(out[0].key, "strong");
+    }
+
+    #[test]
+    fn importance_breaks_relevance_ties() {
+        let a = entry("a", "python script", 0.2, "2024-01-01T00:00:00Z");
+        let b = entry("b", "python script", 0.9, "2024-01-01T00:00:00Z");
+        let terms = vec!["python".to_string(), "script".to_string()];
+        let cfg = RecallConfig::default();
+        let out = rank_and_gate(vec![a, b], &terms, &cfg, 1_704_067_200_000);
+        assert_eq!(out[0].key, "b");
+    }
+
+    #[test]
+    fn recency_decays_with_age() {
+        let recent = recency("2024-01-10T00:00:00Z", 30.0, 1_704_067_200_000);
+        let old = recency("2023-01-10T00:00:00Z", 30.0, 1_704_067_200_000);
+        assert!(recent > old);
+        assert!((0.0..=1.0).contains(&recent));
+    }
+}

src/session/turn_input.rs

@@ -2,7 +2,7 @@ use std::sync::Arc;
 
 use crate::agent::system_prompt::build_runtime_context;
 use crate::bus::ChatMessage;
-use crate::memory::{MemoryCategory, MemoryManager};
+use crate::memory::MemoryManager;
 use crate::work::WorkManager;
 
 /// Immutable context used to assemble provider input for both the initial call
@@ -42,23 +42,20 @@ pub(super) async fn prepare_turn_runtime(
     query: &str,
     system_prompt: String,
 ) -> TurnRuntimeContext {
-    let memory_future = memory_manager.recall(query, 5, Some(MemoryCategory::Knowledge), None);
+    let memory_future = memory_manager.recall_for_context(query);
     let work_future = work_manager.active_plan(session_id);
-    let (memory_result, work_result) = tokio::join!(memory_future, work_future);
+    let (memory_entries, work_result) = tokio::join!(memory_future, work_future);
 
-    let memory_context = match memory_result {
-        Ok(entries) if !entries.is_empty() => Some(
-            entries
+    let memory_context = if memory_entries.is_empty() {
+        None
+    } else {
+        Some(
+            memory_entries
                 .iter()
                 .map(|entry| format!("- {}: {}", entry.key, entry.content))
                 .collect::<Vec<_>>()
                 .join("\n"),
-        ),
-        Err(error) => {
-            tracing::warn!(error = %error, "Failed to fetch memory context");
-            None
-        }
-        _ => None,
+        )
     };
     let work_context = match work_result {
         Ok(Some(plan)) => Some(plan.compact_context()),

src/storage/memory.rs

@@ -1,17 +1,9 @@
 use sqlx::Row;
-use std::sync::OnceLock;
-
-use jieba_rs::Jieba;
 
 use crate::memory::{MemoryCategory, MemoryEntry};
 
 use super::StorageError;
 
-fn jieba() -> &'static Jieba {
-    static INSTANCE: OnceLock<Jieba> = OnceLock::new();
-    INSTANCE.get_or_init(Jieba::new)
-}
-
 impl super::Storage {
     /// List recent memories without requiring a full-text query.
     pub async fn list_memories(
@@ -99,12 +91,27 @@ impl super::Storage {
         session_id: Option<&str>,
         limit: usize,
     ) -> Result<Vec<MemoryEntry>, StorageError> {
-        // Build FTS5 query: segment with jieba, wrap each term in quotes, join with OR
-        let fts_query = jieba()
-            .cut(query, true)
-            .into_iter()
-            .map(|token| token.word)
-            .filter(|word| word.len() > 1 || word.bytes().any(|b| b > 127))
+        let terms = crate::memory::recall::tokenize(query);
+        self.search_memories_by_terms(&terms, category, session_id, limit)
+            .await
+    }
+
+    /// Search memories using pre-tokenized terms (FTS5 with LIKE fallback).
+    /// An empty term list returns no results without issuing a query.
+    pub async fn search_memories_by_terms(
+        &self,
+        terms: &[String],
+        category: Option<&MemoryCategory>,
+        session_id: Option<&str>,
+        limit: usize,
+    ) -> Result<Vec<MemoryEntry>, StorageError> {
+        if terms.is_empty() {
+            return Ok(Vec::new());
+        }
+
+        // Build FTS5 query: wrap each term in quotes, join with OR.
+        let fts_query = terms
+            .iter()
             .map(|word| format!("\"{}\"", word.replace('"', "")))
             .collect::<Vec<_>>()
             .join(" OR ");
@@ -136,15 +143,6 @@ impl super::Storage {
 
         // Fallback to term-based LIKE query if FTS5 returned nothing
         if entries.is_empty() {
-            let terms: Vec<String> = jieba()
-                .cut(query, true)
-                .into_iter()
-                .map(|token| token.word)
-                .filter(|word| word.len() > 1 || word.bytes().any(|b| b > 127))
-                .map(|word| word.replace(['%', '_'], ""))
-                .collect();
-
-            if !terms.is_empty() {
             let like_clauses = terms
                 .iter()
                 .map(|_| "(key LIKE ? OR content LIKE ?)")
@@ -167,8 +165,8 @@ impl super::Storage {
 
             // The only interpolated fragment is a generated sequence of bind placeholders.
             let mut query_builder = sqlx::query(sqlx::AssertSqlSafe(sql));
-                for term in &terms {
-                    let pattern = format!("%{}%", term);
+            for term in terms {
+                let pattern = format!("%{}%", term.replace(['%', '_'], ""));
                 query_builder = query_builder.bind(pattern.clone()).bind(pattern);
             }
             query_builder = query_builder
@@ -181,7 +179,6 @@ impl super::Storage {
             let rows = query_builder.fetch_all(self.pool()).await?;
             entries = parse_memory_rows(&rows)?;
         }
-        }
 
         Ok(entries)
     }
@@ -205,13 +202,7 @@ impl super::Storage {
             .to_rfc3339();
 
         let rows = if let Some(q) = query {
-            let terms: Vec<String> = jieba()
-                .cut(q, true)
-                .into_iter()
-                .map(|token| token.word)
-                .filter(|word| word.len() > 1 || word.bytes().any(|b| b > 127))
-                .map(|word| word.replace(['%', '_'], ""))
-                .collect();
+            let terms: Vec<String> = crate::memory::recall::tokenize(q);
 
             if terms.is_empty() {
                 return Ok(Vec::new());
@@ -241,7 +232,7 @@ impl super::Storage {
             // The only interpolated fragment is a generated sequence of bind placeholders.
             let mut query_builder = sqlx::query(sqlx::AssertSqlSafe(sql));
             for term in &terms {
-                let pattern = format!("%{}%", term);
+                let pattern = format!("%{}%", term.replace(['%', '_'], ""));
                 query_builder = query_builder.bind(pattern.clone()).bind(pattern);
             }
             query_builder = query_builder