feat(memory): add chunk overlap support and clean up per-agent config

- Add ChunkOverlap to MemoryConfig (default 200) for context continuity
  at chunk boundaries during semantic search
- Implement overlap logic in ChunkText() with safety clamp at maxChunkLen/2
- Wire per-agent chunk_len/chunk_overlap overrides at IndexDocument time
  via RunContext, falling back to global PGMemoryConfig defaults
- Add sync.RWMutex to PGMemoryStore for thread-safe config updates
- Persist chunk settings via system_configs (embedding.max_chunk_len,
  embedding.chunk_overlap) with runtime refresh on config change
- Remove dead per-agent embedding_provider/embedding_model fields from
  agent Memory UI (system uses single global provider)
- Show Memory section always-visible in agent config (no toggle needed)
- Add chunk_overlap fields to System Settings modal and Config AI Defaults
- Update i18n (en/vi/zh) for all new and removed keys

cmd/gateway.go

@@ -410,6 +410,12 @@ func runGateway() {
 			}
 			if sysConfigs, err := pgStores.SystemConfigs.List(ctx); err == nil && len(sysConfigs) > 0 {
 				cfg.ApplySystemConfigs(sysConfigs)
+				// Update PGMemoryStore chunk config so new documents use updated settings
+				if mem := cfg.Agents.Defaults.Memory; mem != nil {
+					if pgMem, ok := pgStores.Memory.(*pg.PGMemoryStore); ok {
+						pgMem.UpdateChunkConfig(mem.MaxChunkLen, mem.ChunkOverlap)
+					}
+				}
 				slog.Debug("system_configs refreshed to in-memory config", "keys", len(sysConfigs))
 			}
 		})

cmd/gateway_system_config_sync.go

@@ -87,6 +87,8 @@ func seedConfigForContext(ctx context.Context, sc store.SystemConfigStore, cfg *
 	if m := cfg.Agents.Defaults.Memory; m != nil {
 		set("embedding.provider", m.EmbeddingProvider)
 		set("embedding.model", m.EmbeddingModel)
+		setInt("embedding.max_chunk_len", m.MaxChunkLen)
+		setInt("embedding.chunk_overlap", m.ChunkOverlap)
 	}
 
 	// Agent defaults

internal/config/config.go

@@ -181,6 +181,7 @@ type MemoryConfig struct {
 	EmbeddingAPIBase  string  `json:"embedding_api_base,omitempty"` // custom endpoint URL
 	MaxResults        int     `json:"max_results,omitempty"`        // default 6
 	MaxChunkLen       int     `json:"max_chunk_len,omitempty"`      // default 1000
+	ChunkOverlap      int     `json:"chunk_overlap,omitempty"`      // overlap chars between chunks (default 200)
 	VectorWeight      float64 `json:"vector_weight,omitempty"`      // hybrid search vector weight (default 0.7)
 	TextWeight        float64 `json:"text_weight,omitempty"`        // hybrid search FTS weight (default 0.3)
 	MinScore          float64 `json:"min_score,omitempty"`          // minimum relevance score (default 0.35)

internal/config/config_system.go

@@ -34,6 +34,8 @@ func (c *Config) ApplySystemConfigs(configs map[string]string) {
 	}
 	str("embedding.provider", &c.Agents.Defaults.Memory.EmbeddingProvider)
 	str("embedding.model", &c.Agents.Defaults.Memory.EmbeddingModel)
+	integer("embedding.max_chunk_len", &c.Agents.Defaults.Memory.MaxChunkLen)
+	integer("embedding.chunk_overlap", &c.Agents.Defaults.Memory.ChunkOverlap)
 
 	// Agent defaults
 	str("agent.default_provider", &c.Agents.Defaults.Provider)

internal/memory/embeddings.go

@@ -25,18 +25,31 @@ type TextChunk struct {
 	EndLine   int
 }
 
-// ChunkText splits text into chunks at paragraph boundaries.
+// ChunkText splits text into chunks at paragraph boundaries with optional overlap.
 // Each chunk includes its starting line number in the source file.
-func ChunkText(text string, maxChunkLen int) []TextChunk {
+// When overlap > 0, trailing lines from the previous chunk are prepended to the next chunk
+// so that context at chunk boundaries is preserved for semantic search.
+func ChunkText(text string, maxChunkLen, overlap int) []TextChunk {
 	if maxChunkLen <= 0 {
 		maxChunkLen = 1000
 	}
+	if overlap < 0 {
+		overlap = 0
+	}
+	// Clamp overlap to half the chunk size to prevent infinite loop
+	if overlap >= maxChunkLen/2 {
+		overlap = maxChunkLen / 2
+	}
 
 	lines := strings.Split(text, "\n")
 	var chunks []TextChunk
 	var current strings.Builder
 	startLine := 1
 
+	// overlapLines holds trailing lines from the previous chunk to prepend to the next.
+	var overlapLines []string
+	overlapStartLine := 0
+
 	flush := func(endLine int) {
 		content := strings.TrimSpace(current.String())
 		if content != "" {
@@ -46,8 +59,40 @@ func ChunkText(text string, maxChunkLen int) []TextChunk {
 				EndLine:   endLine,
 			})
 		}
+
+		// Compute overlap lines to carry into the next chunk.
+		overlapLines = nil
+		overlapStartLine = 0
+		if overlap > 0 && endLine > 0 {
+			charCount := 0
+			for j := endLine - 1; j >= startLine-1 && j >= 0; j-- {
+				lineLen := len(lines[j])
+				if charCount+lineLen > overlap {
+					break
+				}
+				charCount += lineLen + 1 // +1 for newline
+				overlapLines = append(overlapLines, lines[j])
+				overlapStartLine = j + 1 // 1-based
+			}
+			// Reverse to restore original order
+			for left, right := 0, len(overlapLines)-1; left < right; left, right = left+1, right-1 {
+				overlapLines[left], overlapLines[right] = overlapLines[right], overlapLines[left]
+			}
+		}
+
 		current.Reset()
-		startLine = endLine + 1
+		// Seed next chunk with overlap content
+		if len(overlapLines) > 0 {
+			startLine = overlapStartLine
+			for k, ol := range overlapLines {
+				if k > 0 {
+					current.WriteString("\n")
+				}
+				current.WriteString(ol)
+			}
+		} else {
+			startLine = endLine + 1
+		}
 	}
 
 	for i, line := range lines {

internal/store/pg/memory_docs.go

@@ -5,6 +5,7 @@ import (
 	"database/sql"
 	"fmt"
 	"log/slog"
+	"sync"
 	"time"
 
 	"github.com/google/uuid"
@@ -17,12 +18,14 @@ import (
 type PGMemoryStore struct {
 	db       *sql.DB
 	provider store.EmbeddingProvider
+	mu       sync.RWMutex   // protects cfg from concurrent read/write
 	cfg      PGMemoryConfig
 }
 
 // PGMemoryConfig configures the PG memory store.
 type PGMemoryConfig struct {
 	MaxChunkLen  int
+	ChunkOverlap int
 	MaxResults   int
 	VectorWeight float64
 	TextWeight   float64
@@ -32,6 +35,7 @@ type PGMemoryConfig struct {
 func DefaultPGMemoryConfig() PGMemoryConfig {
 	return PGMemoryConfig{
 		MaxChunkLen:  1000,
+		ChunkOverlap: 200,
 		MaxResults:   6,
 		VectorWeight: 0.7,
 		TextWeight:   0.3,
@@ -207,8 +211,19 @@ func (s *PGMemoryStore) IndexDocument(ctx context.Context, agentID, userID, path
 	// Delete old chunks
 	s.db.ExecContext(ctx, "DELETE FROM memory_chunks WHERE document_id = $1", docID)
 
+	// Resolve chunk parameters: per-agent override → global default
+	chunkLen, chunkOverlap := s.chunkConfig()
+	if rc := store.RunContextFromCtx(ctx); rc != nil && rc.MemoryCfg != nil {
+		if rc.MemoryCfg.MaxChunkLen > 0 {
+			chunkLen = rc.MemoryCfg.MaxChunkLen
+		}
+		if rc.MemoryCfg.ChunkOverlap > 0 {
+			chunkOverlap = rc.MemoryCfg.ChunkOverlap
+		}
+	}
+
 	// Chunk text
-	chunks := memory.ChunkText(content, s.cfg.MaxChunkLen)
+	chunks := memory.ChunkText(content, chunkLen, chunkOverlap)
 	if len(chunks) == 0 {
 		return nil
 	}
@@ -350,6 +365,25 @@ func (s *PGMemoryStore) SetEmbeddingProvider(provider store.EmbeddingProvider) {
 	s.provider = provider
 }
 
+// UpdateChunkConfig updates chunk splitting parameters at runtime (e.g. after system config change).
+func (s *PGMemoryStore) UpdateChunkConfig(maxChunkLen, chunkOverlap int) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if maxChunkLen > 0 {
+		s.cfg.MaxChunkLen = maxChunkLen
+	}
+	if chunkOverlap >= 0 {
+		s.cfg.ChunkOverlap = chunkOverlap
+	}
+}
+
+// chunkConfig returns a snapshot of the current chunk parameters (thread-safe).
+func (s *PGMemoryStore) chunkConfig() (maxLen, overlap int) {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	return s.cfg.MaxChunkLen, s.cfg.ChunkOverlap
+}
+
 // BackfillEmbeddings finds all chunks without embeddings and generates them.
 // Processes in batches to avoid memory spikes. Safe to call multiple times.
 func (s *PGMemoryStore) BackfillEmbeddings(ctx context.Context) (int, error) {

ui/web/src/components/layout/system-settings-modal.tsx

@@ -64,6 +64,8 @@ interface SystemSettingsModalProps {
 interface InitState {
   embProvider: string;
   embModel: string;
+  embMaxChunkLen: string;
+  embChunkOverlap: string;
   toolStatus: boolean;
   blockReply: boolean;
   intentClassify: boolean;
@@ -76,6 +78,7 @@ interface InitState {
 
 const DEFAULTS: InitState = {
   embProvider: "", embModel: "",
+  embMaxChunkLen: "", embChunkOverlap: "",
   toolStatus: true, blockReply: false, intentClassify: true,
   compProvider: "", compModel: "",
   compThreshold: "", compKeepRecent: "", compMaxTokens: "",
@@ -98,6 +101,8 @@ export function SystemSettingsModal({ open, onOpenChange }: SystemSettingsModalP
   // Embedding
   const [embProvider, setEmbProvider] = useState("");
   const [embModel, setEmbModel] = useState("");
+  const [embMaxChunkLen, setEmbMaxChunkLen] = useState("");
+  const [embChunkOverlap, setEmbChunkOverlap] = useState("");
   const { verifyEmbedding, embVerifying, embResult, resetEmb } = useProviderVerify();
 
   // UX Behavior
@@ -116,6 +121,8 @@ export function SystemSettingsModal({ open, onOpenChange }: SystemSettingsModalP
     const s: InitState = {
       embProvider: configs["embedding.provider"] ?? "",
       embModel: configs["embedding.model"] ?? "",
+      embMaxChunkLen: configs["embedding.max_chunk_len"] ?? "",
+      embChunkOverlap: configs["embedding.chunk_overlap"] ?? "",
       toolStatus: parseBool(configs["gateway.tool_status"], true),
       blockReply: parseBool(configs["gateway.block_reply"], false),
       intentClassify: parseBool(configs["gateway.intent_classify"], true),
@@ -128,6 +135,8 @@ export function SystemSettingsModal({ open, onOpenChange }: SystemSettingsModalP
     setInit(s);
     setEmbProvider(s.embProvider);
     setEmbModel(s.embModel);
+    setEmbMaxChunkLen(s.embMaxChunkLen);
+    setEmbChunkOverlap(s.embChunkOverlap);
     setToolStatus(s.toolStatus);
     setBlockReply(s.blockReply);
     setIntentClassify(s.intentClassify);
@@ -172,6 +181,8 @@ export function SystemSettingsModal({ open, onOpenChange }: SystemSettingsModalP
       const updates: Record<string, string> = {};
       if (embProvider !== init.embProvider) updates["embedding.provider"] = embProvider;
       if (embModel !== init.embModel) updates["embedding.model"] = embModel;
+      if (embMaxChunkLen !== init.embMaxChunkLen) updates["embedding.max_chunk_len"] = embMaxChunkLen;
+      if (embChunkOverlap !== init.embChunkOverlap) updates["embedding.chunk_overlap"] = embChunkOverlap;
       if (toolStatus !== init.toolStatus) updates["gateway.tool_status"] = String(toolStatus);
       if (blockReply !== init.blockReply) updates["gateway.block_reply"] = String(blockReply);
       if (intentClassify !== init.intentClassify) updates["gateway.intent_classify"] = String(intentClassify);
@@ -317,6 +328,19 @@ export function SystemSettingsModal({ open, onOpenChange }: SystemSettingsModalP
                     </span>
                   )}
                 </div>
+
+                <div className="grid grid-cols-1 gap-4 sm:grid-cols-2">
+                  <div className="space-y-1.5">
+                    <Label htmlFor="embMaxChunkLen" className="text-xs">{t("embedding.maxChunkLen")}</Label>
+                    <Input id="embMaxChunkLen" type="number" placeholder="1000" value={embMaxChunkLen} onChange={(e) => setEmbMaxChunkLen(e.target.value)} className="text-base md:text-sm" />
+                    <p className="text-xs text-muted-foreground">{t("embedding.maxChunkLenHint")}</p>
+                  </div>
+                  <div className="space-y-1.5">
+                    <Label htmlFor="embChunkOverlap" className="text-xs">{t("embedding.chunkOverlap")}</Label>
+                    <Input id="embChunkOverlap" type="number" placeholder="200" value={embChunkOverlap} onChange={(e) => setEmbChunkOverlap(e.target.value)} className="text-base md:text-sm" />
+                    <p className="text-xs text-muted-foreground">{t("embedding.chunkOverlapHint")}</p>
+                  </div>
+                </div>
               </CardContent>
             </Card>
 

ui/web/src/i18n/locales/en/agents.json

@@ -383,12 +383,20 @@
     "memory": {
       "title": "Memory",
       "description": "Semantic memory search and embedding configuration",
-      "embeddingProvider": "Embedding Provider",
-      "embeddingModel": "Embedding Model",
-      "maxResults": "Max Results",
+      "enabled": "Enabled",
+      "enabledTip": "Enable or disable the memory system for this agent.",
       "maxChunkLen": "Max Chunk Length",
+      "maxChunkLenTip": "Maximum characters per memory chunk. Leave empty for system default.",
+      "chunkOverlap": "Chunk Overlap",
+      "chunkOverlapTip": "Overlap characters between chunks for context continuity. Leave empty for system default.",
+      "maxResults": "Max Results",
+      "maxResultsTip": "Maximum memory entries returned per search query.",
+      "minScore": "Min Score",
+      "minScoreTip": "Minimum relevance score for search results.",
       "vectorWeight": "Vector Weight",
-      "textWeight": "Text Weight"
+      "vectorWeightTip": "Weight for semantic similarity in hybrid search.",
+      "textWeight": "Text Weight",
+      "textWeightTip": "Weight for keyword similarity in hybrid search."
     },
     "thinking": {
       "title": "Extended Thinking",

ui/web/src/i18n/locales/en/config.json

@@ -87,6 +87,10 @@
   "agents.memory.embeddingModelTip": "Embedding model name. Leave empty to use the provider default.",
   "agents.memory.maxResults": "Max Results",
   "agents.memory.maxResultsTip": "Maximum number of memory entries returned per recall query.",
+  "agents.memory.maxChunkLen": "Max Chunk Length",
+  "agents.memory.maxChunkLenTip": "Maximum characters per memory chunk before splitting.",
+  "agents.memory.chunkOverlap": "Chunk Overlap",
+  "agents.memory.chunkOverlapTip": "Overlap characters between chunks for context continuity.",
   "agents.memory.minScore": "Min Score",
   "agents.memory.minScoreTip": "Minimum similarity score (0-1) for a memory entry to be included in results.",
 

ui/web/src/i18n/locales/en/system-settings.json

@@ -21,7 +21,11 @@
     "verifyRequired": "Verify the embedding model before saving.",
     "verifyFailed": "Verification failed",
     "dimensions": "{{count}} dimensions",
-    "dimensionsMismatch": "{{count}} dimensions — does not match required 1536"
+    "dimensionsMismatch": "{{count}} dimensions — does not match required 1536",
+    "maxChunkLen": "Max Chunk Length",
+    "maxChunkLenHint": "Maximum characters per memory chunk before splitting.",
+    "chunkOverlap": "Chunk Overlap",
+    "chunkOverlapHint": "Overlap characters between consecutive chunks for context continuity."
   },
   "ux": {
     "title": "UX Behavior",

ui/web/src/i18n/locales/vi/agents.json

@@ -383,12 +383,20 @@
     "memory": {
       "title": "Bộ nhớ",
       "description": "Tìm kiếm bộ nhớ ngữ nghĩa và cấu hình nhúng",
-      "embeddingProvider": "Provider nhúng",
-      "embeddingModel": "Model nhúng",
-      "maxResults": "Kết quả tối đa",
+      "enabled": "Đã bật",
+      "enabledTip": "Bật/tắt hệ thống bộ nhớ cho agent này.",
       "maxChunkLen": "Độ dài chunk tối đa",
+      "maxChunkLenTip": "Số ký tự tối đa mỗi chunk. Để trống dùng mặc định hệ thống.",
+      "chunkOverlap": "Chunk Overlap",
+      "chunkOverlapTip": "Số ký tự chồng lấp giữa các chunk để giữ ngữ cảnh. Để trống dùng mặc định.",
+      "maxResults": "Kết quả tối đa",
+      "maxResultsTip": "Số kết quả tối đa trả về mỗi truy vấn tìm kiếm.",
+      "minScore": "Điểm tối thiểu",
+      "minScoreTip": "Ngưỡng điểm tối thiểu cho kết quả tìm kiếm.",
       "vectorWeight": "Trọng số vector",
-      "textWeight": "Trọng số văn bản"
+      "vectorWeightTip": "Trọng số tương đồng ngữ nghĩa trong tìm kiếm hỗn hợp.",
+      "textWeight": "Trọng số văn bản",
+      "textWeightTip": "Trọng số từ khóa trong tìm kiếm hỗn hợp."
     },
     "thinking": {
       "title": "Suy nghĩ mở rộng",

ui/web/src/i18n/locales/vi/config.json

@@ -87,6 +87,10 @@
   "agents.memory.embeddingModelTip": "Tên model nhúng. Để trống để dùng mặc định của provider.",
   "agents.memory.maxResults": "Kết quả tối đa",
   "agents.memory.maxResultsTip": "Số mục bộ nhớ tối đa trả về cho mỗi truy vấn.",
+  "agents.memory.maxChunkLen": "Độ dài chunk tối đa",
+  "agents.memory.maxChunkLenTip": "Số ký tự tối đa mỗi chunk trước khi chia nhỏ.",
+  "agents.memory.chunkOverlap": "Chunk Overlap",
+  "agents.memory.chunkOverlapTip": "Số ký tự chồng lấp giữa các chunk để giữ ngữ cảnh.",
   "agents.memory.minScore": "Điểm tối thiểu",
   "agents.memory.minScoreTip": "Điểm tương đồng tối thiểu (0-1) để đưa mục bộ nhớ vào kết quả.",
 

ui/web/src/i18n/locales/vi/system-settings.json

@@ -21,7 +21,11 @@
     "verifyRequired": "Kiểm tra model embedding trước khi lưu.",
     "verifyFailed": "Kiểm tra thất bại",
     "dimensions": "{{count}} chiều",
-    "dimensionsMismatch": "{{count}} chiều — không khớp yêu cầu 1536"
+    "dimensionsMismatch": "{{count}} chiều — không khớp yêu cầu 1536",
+    "maxChunkLen": "Độ dài chunk tối đa",
+    "maxChunkLenHint": "Số ký tự tối đa mỗi chunk trước khi chia nhỏ.",
+    "chunkOverlap": "Chunk Overlap",
+    "chunkOverlapHint": "Số ký tự chồng lấp giữa các chunk liên tiếp để giữ ngữ cảnh."
   },
   "ux": {
     "title": "Hành vi UX",

ui/web/src/i18n/locales/zh/agents.json

@@ -383,12 +383,20 @@
     "memory": {
       "title": "记忆",
       "description": "语义记忆搜索和嵌入配置",
-      "embeddingProvider": "嵌入Provider",
-      "embeddingModel": "嵌入模型",
-      "maxResults": "最大结果数",
+      "enabled": "已启用",
+      "enabledTip": "启用或禁用此Agent的记忆系统。",
       "maxChunkLen": "最大块长度",
+      "maxChunkLenTip": "每个记忆块的最大字符数。留空使用系统默认值。",
+      "chunkOverlap": "块重叠",
+      "chunkOverlapTip": "连续块之间的重叠字符数。留空使用系统默认值。",
+      "maxResults": "最大结果数",
+      "maxResultsTip": "每次搜索返回的最大记忆条目数。",
+      "minScore": "最低分数",
+      "minScoreTip": "搜索结果的最低相关性分数阈值。",
       "vectorWeight": "向量权重",
-      "textWeight": "文本权重"
+      "vectorWeightTip": "混合搜索中语义相似度的权重。",
+      "textWeight": "文本权重",
+      "textWeightTip": "混合搜索中关键词相似度的权重。"
     },
     "thinking": {
       "title": "扩展思考",

ui/web/src/i18n/locales/zh/config.json

@@ -87,6 +87,10 @@
   "agents.memory.embeddingModelTip": "嵌入模型名称。留空则使用提供商默认值。",
   "agents.memory.maxResults": "最大结果数",
   "agents.memory.maxResultsTip": "每次召回查询返回的最大记忆条目数。",
+  "agents.memory.maxChunkLen": "最大块长度",
+  "agents.memory.maxChunkLenTip": "每个记忆块分割前的最大字符数。",
+  "agents.memory.chunkOverlap": "块重叠",
+  "agents.memory.chunkOverlapTip": "连续块之间的重叠字符数，用于保持上下文连续性。",
   "agents.memory.minScore": "最低分数",
   "agents.memory.minScoreTip": "记忆条目被纳入结果的最低相似度分数（0-1）。",
 

ui/web/src/i18n/locales/zh/system-settings.json

@@ -21,7 +21,11 @@
     "verifyRequired": "保存前请先验证Embedding模型。",
     "verifyFailed": "验证失败",
     "dimensions": "{{count}}维",
-    "dimensionsMismatch": "{{count}}维 — 不符合要求的1536"
+    "dimensionsMismatch": "{{count}}维 — 不符合要求的1536",
+    "maxChunkLen": "最大块长度",
+    "maxChunkLenHint": "每个记忆块分割前的最大字符数。",
+    "chunkOverlap": "块重叠",
+    "chunkOverlapHint": "连续块之间的重叠字符数，用于保持上下文连续性。"
   },
   "ux": {
     "title": "UX行为",

ui/web/src/pages/agents/agent-detail/agent-overview-tab.tsx

@@ -11,7 +11,6 @@ import { EvolutionSection } from "./overview-sections/evolution-section";
 import { CapabilitiesSection } from "./overview-sections/capabilities-section";
 import { HeartbeatCard } from "./overview-sections/heartbeat-card";
 import { MemorySection } from "./config-sections";
-import { ConfigGroupHeader } from "@/components/shared/config-group-header";
 import type { UseAgentHeartbeatReturn } from "../hooks/use-agent-heartbeat";
 
 interface AgentOverviewTabProps {
@@ -47,8 +46,7 @@ export function AgentOverviewTab({ agent, onUpdate, heartbeat }: AgentOverviewTa
     typeof otherCfg.skill_nudge_interval === "number" ? otherCfg.skill_nudge_interval : 15,
   );
 
-  // Memory (now standalone section)
-  const [memEnabled, setMemEnabled] = useState(agent.memory_config != null);
+  // Memory (always shown — per-agent overrides, empty = use system defaults)
   const [mem, setMem] = useState<MemoryConfig>(agent.memory_config ?? {});
 
   // Capabilities (subagents + tool policy)
@@ -83,7 +81,7 @@ export function AgentOverviewTab({ agent, onUpdate, heartbeat }: AgentOverviewTa
         is_default: isDefault,
         other_config: updatedOtherConfig,
         budget_monthly_cents: budgetCents,
-        memory_config: memEnabled ? mem : null,
+        memory_config: mem,
         subagents_config: subEnabled ? sub : null,
         tools_config: toolsEnabled
           ? { profile: tools.profile, allow: tools.allow, deny: tools.deny, alsoAllow: tools.alsoAllow, byProvider: tools.byProvider }
@@ -128,19 +126,11 @@ export function AgentOverviewTab({ agent, onUpdate, heartbeat }: AgentOverviewTa
         onSaveBlockedChange={setLlmSaveBlocked}
       />
 
-      {/* Memory — standalone section (was inside Capabilities) */}
-      <section className="space-y-4">
-        <ConfigGroupHeader
-          title={t("configSections.memory.title")}
-          description={t("configSections.memory.description")}
-        />
-        <MemorySection
-          enabled={memEnabled}
-          value={mem}
-          onToggle={(v) => { setMemEnabled(v); if (!v) setMem({}); }}
-          onChange={setMem}
-        />
-      </section>
+      {/* Memory — always visible, per-agent overrides */}
+      <MemorySection
+        value={mem}
+        onChange={setMem}
+      />
 
       <HeartbeatCard heartbeat={heartbeat} />
 

ui/web/src/pages/agents/agent-detail/config-sections/memory-section.tsx

@@ -2,89 +2,100 @@ import { useTranslation } from "react-i18next";
 import { Input } from "@/components/ui/input";
 import { Switch } from "@/components/ui/switch";
 import type { MemoryConfig } from "@/types/agent";
-import { ConfigSection, InfoLabel, numOrUndef } from "./config-section";
-import { ProviderModelSelect } from "@/components/shared/provider-model-select";
+import { InfoLabel, numOrUndef } from "./config-section";
 
 interface MemorySectionProps {
-  enabled: boolean;
   value: MemoryConfig;
-  onToggle: (v: boolean) => void;
   onChange: (v: MemoryConfig) => void;
 }
 
-export function MemorySection({ enabled, value, onToggle, onChange }: MemorySectionProps) {
+export function MemorySection({ value, onChange }: MemorySectionProps) {
   const { t } = useTranslation("agents");
   const s = "configSections.memory";
   return (
-    <ConfigSection
-      title={t(`${s}.title`)}
-      description={t(`${s}.description`)}
-      enabled={enabled}
-      onToggle={onToggle}
-    >
-      <div className="flex items-center gap-2">
-        <Switch
-          checked={value.enabled ?? true}
-          onCheckedChange={(v) => onChange({ ...value, enabled: v })}
-        />
-        <InfoLabel tip="Enable or disable the memory system for this agent. When enabled, the agent can store and recall information across sessions.">{t("configSections.contextPruning.enabled")}</InfoLabel>
+    <section className="space-y-3">
+      <div>
+        <h3 className="text-sm font-medium">{t(`${s}.title`)}</h3>
+        <p className="text-xs text-muted-foreground">{t(`${s}.description`)}</p>
       </div>
-      <ProviderModelSelect
-        provider={value.embedding_provider ?? ""}
-        onProviderChange={(v) => onChange({ ...value, embedding_provider: v || undefined })}
-        model={value.embedding_model ?? ""}
-        onModelChange={(v) => onChange({ ...value, embedding_model: v || undefined })}
-        providerLabel={t(`${s}.embeddingProvider`)}
-        modelLabel={t(`${s}.embeddingModel`)}
-        providerTip="LLM provider used for generating text embeddings. Leave empty to auto-detect. Only providers with embedding enabled are shown."
-        modelTip="Embedding model name (e.g. text-embedding-3-small). Must be supported by the provider."
-        providerPlaceholder="(auto)"
-        modelPlaceholder="text-embedding-3-small"
-        allowEmpty
-        filterEmbedding
-      />
-      <div className="grid grid-cols-1 gap-4 sm:grid-cols-2">
-        <div className="space-y-2">
-          <InfoLabel tip="Maximum number of memory entries returned per search query.">{t(`${s}.maxResults`)}</InfoLabel>
-          <Input
-            type="number"
-            placeholder="6"
-            value={value.max_results ?? ""}
-            onChange={(e) => onChange({ ...value, max_results: numOrUndef(e.target.value) })}
+      <div className="rounded-lg border p-3 space-y-4 sm:p-4">
+        <div className="flex items-center gap-2">
+          <Switch
+            checked={value.enabled ?? true}
+            onCheckedChange={(v) => onChange({ ...value, enabled: v })}
           />
+          <InfoLabel tip={t(`${s}.enabledTip`)}>{t(`${s}.enabled`)}</InfoLabel>
         </div>
-        <div className="space-y-2">
-          <InfoLabel tip="Maximum character length for each memory chunk. Longer content is split into smaller chunks before storing.">{t(`${s}.maxChunkLen`)}</InfoLabel>
-          <Input
-            type="number"
-            placeholder="1000"
-            value={value.max_chunk_len ?? ""}
-            onChange={(e) => onChange({ ...value, max_chunk_len: numOrUndef(e.target.value) })}
-          />
+        <div className="grid grid-cols-1 gap-4 sm:grid-cols-2">
+          <div className="space-y-2">
+            <InfoLabel tip={t(`${s}.maxChunkLenTip`)}>{t(`${s}.maxChunkLen`)}</InfoLabel>
+            <Input
+              type="number"
+              placeholder="1000"
+              value={value.max_chunk_len ?? ""}
+              onChange={(e) => onChange({ ...value, max_chunk_len: numOrUndef(e.target.value) })}
+              className="text-base md:text-sm"
+            />
+          </div>
+          <div className="space-y-2">
+            <InfoLabel tip={t(`${s}.chunkOverlapTip`)}>{t(`${s}.chunkOverlap`)}</InfoLabel>
+            <Input
+              type="number"
+              placeholder="200"
+              value={value.chunk_overlap ?? ""}
+              onChange={(e) => onChange({ ...value, chunk_overlap: numOrUndef(e.target.value) })}
+              className="text-base md:text-sm"
+            />
+          </div>
+        </div>
+        <div className="grid grid-cols-1 gap-4 sm:grid-cols-2">
+          <div className="space-y-2">
+            <InfoLabel tip={t(`${s}.maxResultsTip`)}>{t(`${s}.maxResults`)}</InfoLabel>
+            <Input
+              type="number"
+              placeholder="6"
+              value={value.max_results ?? ""}
+              onChange={(e) => onChange({ ...value, max_results: numOrUndef(e.target.value) })}
+              className="text-base md:text-sm"
+            />
+          </div>
+          <div className="space-y-2">
+            <InfoLabel tip={t(`${s}.minScoreTip`)}>{t(`${s}.minScore`)}</InfoLabel>
+            <Input
+              type="number"
+              step="0.01"
+              placeholder="0.35"
+              value={value.min_score ?? ""}
+              onChange={(e) => onChange({ ...value, min_score: numOrUndef(e.target.value) })}
+              className="text-base md:text-sm"
+            />
+          </div>
+        </div>
+        <div className="grid grid-cols-1 gap-4 sm:grid-cols-2">
+          <div className="space-y-2">
+            <InfoLabel tip={t(`${s}.vectorWeightTip`)}>{t(`${s}.vectorWeight`)}</InfoLabel>
+            <Input
+              type="number"
+              step="0.1"
+              placeholder="0.7"
+              value={value.vector_weight ?? ""}
+              onChange={(e) => onChange({ ...value, vector_weight: numOrUndef(e.target.value) })}
+              className="text-base md:text-sm"
+            />
+          </div>
+          <div className="space-y-2">
+            <InfoLabel tip={t(`${s}.textWeightTip`)}>{t(`${s}.textWeight`)}</InfoLabel>
+            <Input
+              type="number"
+              step="0.1"
+              placeholder="0.3"
+              value={value.text_weight ?? ""}
+              onChange={(e) => onChange({ ...value, text_weight: numOrUndef(e.target.value) })}
+              className="text-base md:text-sm"
+            />
+          </div>
         </div>
       </div>
-      <div className="grid grid-cols-1 gap-4 sm:grid-cols-2">
-        <div className="space-y-2">
-          <InfoLabel tip="Weight for vector (semantic) similarity in hybrid search scoring. Higher values prioritize meaning over keywords.">{t(`${s}.vectorWeight`)}</InfoLabel>
-          <Input
-            type="number"
-            step="0.1"
-            placeholder="0.7"
-            value={value.vector_weight ?? ""}
-            onChange={(e) => onChange({ ...value, vector_weight: numOrUndef(e.target.value) })}
-          />
-        </div>
-        <div className="space-y-2">
-          <InfoLabel tip="Weight for text (keyword/BM25) similarity in hybrid search scoring. Higher values prioritize exact keyword matches.">{t(`${s}.textWeight`)}</InfoLabel>
-          <Input
-            type="number"
-            step="0.1"
-            placeholder="0.3"
-            value={value.text_weight ?? ""}
-            onChange={(e) => onChange({ ...value, text_weight: numOrUndef(e.target.value) })}
-          />
-        </div>
-      </div>
-    </ConfigSection>
+    </section>
   );
 }

ui/web/src/pages/config/sections/ai-defaults-section.tsx

@@ -180,6 +180,10 @@ export function AiDefaultsSection({ data, onSave, saving }: Props) {
             <Field label={t("agents.memory.maxResults")} tip={t("agents.memory.maxResultsTip")} type="number" value={memory.max_results} onChange={(v) => updateNested("memory", { max_results: Number(v) })} placeholder="6" />
             <Field label={t("agents.memory.minScore")} tip={t("agents.memory.minScoreTip")} type="number" step="0.01" value={memory.min_score} onChange={(v) => updateNested("memory", { min_score: Number(v) })} placeholder="0.35" />
           </div>
+          <div className="grid grid-cols-1 gap-4 sm:grid-cols-2">
+            <Field label={t("agents.memory.maxChunkLen")} tip={t("agents.memory.maxChunkLenTip")} type="number" value={memory.max_chunk_len} onChange={(v) => updateNested("memory", { max_chunk_len: Number(v) })} placeholder="1000" />
+            <Field label={t("agents.memory.chunkOverlap")} tip={t("agents.memory.chunkOverlapTip")} type="number" value={memory.chunk_overlap} onChange={(v) => updateNested("memory", { chunk_overlap: Number(v) })} placeholder="200" />
+          </div>
         </SubSection>
 
         <SubSection

ui/web/src/types/agent.ts

@@ -64,8 +64,10 @@ export interface MemoryConfig {
   embedding_model?: string;
   max_results?: number;
   max_chunk_len?: number;
+  chunk_overlap?: number;
   vector_weight?: number;
   text_weight?: number;
+  min_score?: number;
 }
 
 export interface WorkspaceSharingConfig {