feat: introduce MemoryPromptConfig for customizable memory prompts

docs/ar/concepts/memory.mdx

@@ -157,6 +157,43 @@ class ResearchFlow(Flow):
 
 انظر [وثائق التدفقات](/concepts/flows) لمزيد من المعلومات حول الذاكرة في التدفقات.
 
+## تخصيص مطالبات الذاكرة (`MemoryPromptConfig`)
+
+يمكنك استبدال تعليمات نموذج اللغة في كل خطوة من تحليل الذاكرة (نفس فكرة ضبط مطالبات التخطيط). مرّر كائن `MemoryPromptConfig` كوسيط `memory_prompt` إلى `Memory`. عيّن الحقول التي تحتاجها فقط؛ تبقى الخطوات الأخرى على القيم الافتراضية المضمّنة في `translations/en.json` تحت المفتاح `memory` (أسماء الحقول تطابق مفاتيح JSON).
+
+```python
+from crewai import Memory, MemoryPromptConfig
+
+memory = Memory(
+    llm="gpt-4o-mini",
+    memory_prompt=MemoryPromptConfig(
+        save_system="...",   # اختياري
+        query_user="...",    # اختياري
+    ),
+)
+```
+
+يمكنك أيضًا تمرير `memory_prompt` إلى دوال مساعدة في `crewai.memory.analyze` (مثل `extract_memories_from_content`) عند استدعائها مباشرة.
+
+### تأثير كل زوج من المطالبات
+
+| الحقول | متى يعمل | ماذا يؤثر |
+| --- | --- | --- |
+| `save_system` / `save_user` | عند الحفظ (`analyze_for_save`) | `suggested_scope` و`categories` و`importance` و`extracted_metadata` المستنتجة قبل التخزين والتضمين. |
+| `query_system` / `query_user` | عند تحليل استعلام الاسترجاع (`analyze_query`) | `keywords` و`suggested_scopes` و`complexity` و`recall_queries` و`time_filter`، ما يوجّه البحث المتجهي وعمق الاسترجاع. |
+| `extract_memories_system` / `extract_memories_user` | `extract_memories_from_content` / `Memory.extract_memories` | كيفية تقسيم النص الخام إلى جمل ذاكرة منفصلة (لا يزال التخزين عبر `remember()`). |
+| `consolidation_system` / `consolidation_user` | عندما يكون المحتوى الجديد قريبًا دلاليًا من سجلات موجودة (`analyze_for_consolidation`) | الإبقاء على الصفوف أو تحديثها أو حذفها، وما إذا كان يُدرج المحتوى الجديد كذاكرة مستقلة. |
+
+### العناصر النائبة (placeholders)
+
+سلاسل **النظام (system)** تُرسل كما هي. سلاسل **المستخدم (user)** تُملأ بـ `str.format` في بايثون. يجب أن تتضمن قوالب المستخدم المخصصة نفس أسماء العناصر النائبة الافتراضية وإلا يفشل التنسيق.
+
+| حقل المستخدم | عناصر نائبة مطلوبة |
+| --- | --- |
+| `save_user` | `{content}`، `{existing_scopes}`، `{existing_categories}` |
+| `query_user` | `{query}`، `{available_scopes}`، `{scope_desc}` |
+| `extract_memories_user` | `{content}` |
+| `consolidation_user` | `{new_content}`، `{records_summary}` |
 
 ## النطاقات الهرمية
 

docs/en/concepts/memory.mdx

@@ -157,6 +157,43 @@ class ResearchFlow(Flow):
 
 See the [Flows documentation](/concepts/flows) for more on memory in Flows.
 
+## Customizing memory prompts (`MemoryPromptConfig`)
+
+Override the LLM instructions used at each memory analysis step (same idea as tuning planning prompts). Pass a `MemoryPromptConfig` as `memory_prompt` on `Memory`. Only set the fields you need; every other step keeps the bundled defaults from the library’s `translations/en.json` under the `memory` key (field names match those JSON keys).
+
+```python
+from crewai import Memory, MemoryPromptConfig
+
+memory = Memory(
+    llm="gpt-4o-mini",
+    memory_prompt=MemoryPromptConfig(
+        save_system="...",   # optional
+        query_user="...",    # optional
+    ),
+)
+```
+
+You can also pass `memory_prompt` into helpers in `crewai.memory.analyze` (for example `extract_memories_from_content`) when you call them directly.
+
+### What each prompt pair affects
+
+| Fields | When it runs | What it influences |
+| --- | --- | --- |
+| `save_system` / `save_user` | Saving (`analyze_for_save`) | Inferred `suggested_scope`, `categories`, `importance`, and `extracted_metadata` before storage and embedding. |
+| `query_system` / `query_user` | Recall query analysis (`analyze_query`) | `keywords`, `suggested_scopes`, `complexity`, `recall_queries`, and `time_filter`, which steer vector search and how deep recall goes. |
+| `extract_memories_system` / `extract_memories_user` | `extract_memories_from_content` / `Memory.extract_memories` | How raw text is split into discrete memory strings (persistence is still via `remember()`). |
+| `consolidation_system` / `consolidation_user` | When new content is similar to existing records (`analyze_for_consolidation`) | Whether to keep, update, or delete existing rows and whether to insert the new content as its own memory. |
+
+### Placeholders
+
+**System** strings are sent as-is. **User** strings are filled with Python’s `str.format`. Custom user templates must include the same placeholder names as the defaults or formatting will raise.
+
+| User field | Required placeholders |
+| --- | --- |
+| `save_user` | `{content}`, `{existing_scopes}`, `{existing_categories}` |
+| `query_user` | `{query}`, `{available_scopes}`, `{scope_desc}` |
+| `extract_memories_user` | `{content}` |
+| `consolidation_user` | `{new_content}`, `{records_summary}` |
 
 ## Hierarchical Scopes
 

docs/ko/concepts/memory.mdx

@@ -157,6 +157,43 @@ class ResearchFlow(Flow):
 
 Flow에서의 메모리에 대한 자세한 내용은 [Flows 문서](/concepts/flows)를 참조하세요.
 
+## 메모리 프롬프트 사용자 지정 (`MemoryPromptConfig`)
+
+메모리 분석 단계마다 사용되는 LLM 지시문을 덮어쓸 수 있습니다(플래닝 프롬프트를 조정하는 것과 같은 개념). `Memory`의 `memory_prompt`에 `MemoryPromptConfig`를 넘깁니다. 필요한 필드만 설정하면 되고, 나머지 단계는 라이브러리 번들 기본값(`translations/en.json`의 `memory` 키; 필드 이름이 해당 JSON 키와 일치)을 그대로 씁니다.
+
+```python
+from crewai import Memory, MemoryPromptConfig
+
+memory = Memory(
+    llm="gpt-4o-mini",
+    memory_prompt=MemoryPromptConfig(
+        save_system="...",   # 선택
+        query_user="...",    # 선택
+    ),
+)
+```
+
+`crewai.memory.analyze`의 헬퍼(예: `extract_memories_from_content`)를 직접 호출할 때도 `memory_prompt`를 넘길 수 있습니다.
+
+### 프롬프트 쌍별 역할
+
+| 필드 | 실행 시점 | 영향 |
+| --- | --- | --- |
+| `save_system` / `save_user` | 저장 시 (`analyze_for_save`) | 저장·임베딩 전에 추론되는 `suggested_scope`, `categories`, `importance`, `extracted_metadata`. |
+| `query_system` / `query_user` | 리콜 시 쿼리 분석 (`analyze_query`) | `keywords`, `suggested_scopes`, `complexity`, `recall_queries`, `time_filter` — 벡터 검색과 리콜 탐색 깊이에 영향. |
+| `extract_memories_system` / `extract_memories_user` | `extract_memories_from_content` / `Memory.extract_memories` | 긴 텍스트를 개별 메모리 문자열로 나누는 방식(저장은 여전히 `remember()`). |
+| `consolidation_system` / `consolidation_user` | 신규 콘텐츠가 기존 레코드와 유사할 때 (`analyze_for_consolidation`) | 기존 행 유지·갱신·삭제 및 신규 콘텐츠를 별도 메모리로 넣을지 여부. |
+
+### 플레이스홀더
+
+**system** 문자열은 그대로 전송됩니다. **user** 문자열은 Python `str.format`으로 채워집니다. 사용자 정의 user 템플릿에는 기본값과 동일한 플레이스홀더 이름이 포함되어야 하며, 그렇지 않으면 포맷 단계에서 오류가 납니다.
+
+| User 필드 | 필수 플레이스홀더 |
+| --- | --- |
+| `save_user` | `{content}`, `{existing_scopes}`, `{existing_categories}` |
+| `query_user` | `{query}`, `{available_scopes}`, `{scope_desc}` |
+| `extract_memories_user` | `{content}` |
+| `consolidation_user` | `{new_content}`, `{records_summary}` |
 
 ## 계층적 범위(Scopes)
 

docs/pt-BR/concepts/memory.mdx

@@ -157,6 +157,43 @@ class ResearchFlow(Flow):
 
 Veja a [documentação de Flows](/concepts/flows) para mais informações sobre memória em Flows.
 
+## Personalizando prompts de memória (`MemoryPromptConfig`)
+
+Substitua as instruções do LLM usadas em cada etapa de análise de memória (mesma ideia que ajustar prompts de planejamento). Passe um `MemoryPromptConfig` como `memory_prompt` em `Memory`. Defina apenas os campos necessários; nas demais etapas permanecem os padrões embutidos do `translations/en.json` da biblioteca, na chave `memory` (os nomes dos campos correspondem às chaves JSON).
+
+```python
+from crewai import Memory, MemoryPromptConfig
+
+memory = Memory(
+    llm="gpt-4o-mini",
+    memory_prompt=MemoryPromptConfig(
+        save_system="...",   # opcional
+        query_user="...",    # opcional
+    ),
+)
+```
+
+Você também pode passar `memory_prompt` para funções auxiliares em `crewai.memory.analyze` (por exemplo `extract_memories_from_content`) quando chamá-las diretamente.
+
+### O que cada par de prompts afeta
+
+| Campos | Quando roda | O que influencia |
+| --- | --- | --- |
+| `save_system` / `save_user` | Ao salvar (`analyze_for_save`) | `suggested_scope`, `categories`, `importance` e `extracted_metadata` inferidos antes do armazenamento e do embedding. |
+| `query_system` / `query_user` | Análise da consulta no recall (`analyze_query`) | `keywords`, `suggested_scopes`, `complexity`, `recall_queries` e `time_filter`, que orientam a busca vetorial e a profundidade do recall. |
+| `extract_memories_system` / `extract_memories_user` | `extract_memories_from_content` / `Memory.extract_memories` | Como o texto bruto é dividido em memórias atômicas (a persistência continua sendo via `remember()`). |
+| `consolidation_system` / `consolidation_user` | Quando o novo conteúdo é semelhante a registros existentes (`analyze_for_consolidation`) | Manter, atualizar ou excluir linhas existentes e se o novo conteúdo entra como memória própria. |
+
+### Placeholders
+
+Strings de **system** são enviadas como estão. Strings de **user** são preenchidas com `str.format` do Python. Templates de user personalizados devem incluir os mesmos nomes de placeholder dos padrões; caso contrário, a formatação falha.
+
+| Campo user | Placeholders obrigatórios |
+| --- | --- |
+| `save_user` | `{content}`, `{existing_scopes}`, `{existing_categories}` |
+| `query_user` | `{query}`, `{available_scopes}`, `{scope_desc}` |
+| `extract_memories_user` | `{content}` |
+| `consolidation_user` | `{new_content}`, `{records_summary}` |
 
 ## Escopos Hierárquicos
 

lib/crewai/src/crewai/__init__.py

@@ -52,6 +52,7 @@ def filtered_warn(
 
 _LAZY_IMPORTS: dict[str, tuple[str, str]] = {
     "Memory": ("crewai.memory.unified_memory", "Memory"),
+    "MemoryPromptConfig": ("crewai.memory.types", "MemoryPromptConfig"),
 }
 
 
@@ -196,6 +197,7 @@ def __getattr__(name: str) -> Any:
     "Knowledge",
     "LLMGuardrail",
     "Memory",
+    "MemoryPromptConfig",
     "PlanningConfig",
     "Process",
     "RuntimeState",

lib/crewai/src/crewai/memory/analyze.py

@@ -8,7 +8,7 @@
 
 from pydantic import BaseModel, ConfigDict, Field
 
-from crewai.memory.types import MemoryRecord, ScopeInfo
+from crewai.memory.types import MemoryPromptConfig, MemoryRecord, ScopeInfo
 from crewai.utilities.i18n import I18N_DEFAULT
 
 
@@ -140,19 +140,23 @@ class ConsolidationPlan(BaseModel):
     )
 
 
-def _get_prompt(key: str) -> str:
-    """Retrieve a memory prompt from the i18n translations.
-
-    Args:
-        key: The prompt key under the "memory" section.
-
-    Returns:
-        The prompt string.
-    """
+def _memory_prompt_line(
+    memory_prompt: MemoryPromptConfig | None,
+    key: str,
+) -> str:
+    """Resolve one memory prompt: override string or bundled translation."""
+    if memory_prompt is not None:
+        raw = getattr(memory_prompt, key, None)
+        if isinstance(raw, str) and raw.strip():
+            return raw
     return I18N_DEFAULT.memory(key)
 
 
-def extract_memories_from_content(content: str, llm: Any) -> list[str]:
+def extract_memories_from_content(
+    content: str,
+    llm: Any,
+    memory_prompt: MemoryPromptConfig | None = None,
+) -> list[str]:
     """Use the LLM to extract discrete memory statements from raw content.
 
     This is a pure helper: it does NOT store anything. Callers should call
@@ -164,15 +168,21 @@ def extract_memories_from_content(content: str, llm: Any) -> list[str]:
     Args:
         content: Raw text (e.g. task description + result dump).
         llm: The LLM instance to use.
+        memory_prompt: Optional per-step prompt strings (see ``MemoryPromptConfig``).
 
     Returns:
         List of short, self-contained memory statements (or [content] on failure).
     """
     if not (content or "").strip():
         return []
-    user = _get_prompt("extract_memories_user").format(content=content)
+    user = _memory_prompt_line(memory_prompt, "extract_memories_user").format(
+        content=content
+    )
     messages = [
-        {"role": "system", "content": _get_prompt("extract_memories_system")},
+        {
+            "role": "system",
+            "content": _memory_prompt_line(memory_prompt, "extract_memories_system"),
+        },
         {"role": "user", "content": user},
     ]
     try:
@@ -202,6 +212,7 @@ def analyze_query(
     available_scopes: list[str],
     scope_info: ScopeInfo | None,
     llm: Any,
+    memory_prompt: MemoryPromptConfig | None = None,
 ) -> QueryAnalysis:
     """Use the LLM to analyze a recall query.
 
@@ -212,20 +223,24 @@ def analyze_query(
         available_scopes: Scope paths that exist in the store.
         scope_info: Optional info about the current scope.
         llm: The LLM instance to use.
+        memory_prompt: Optional per-step prompt strings.
 
     Returns:
         QueryAnalysis with keywords, suggested_scopes, complexity, recall_queries, time_filter.
     """
     scope_desc = ""
     if scope_info:
         scope_desc = f"Current scope has {scope_info.record_count} records, categories: {scope_info.categories}"
-    user = _get_prompt("query_user").format(
+    user = _memory_prompt_line(memory_prompt, "query_user").format(
         query=query,
         available_scopes=available_scopes or ["/"],
         scope_desc=scope_desc,
     )
     messages = [
-        {"role": "system", "content": _get_prompt("query_system")},
+        {
+            "role": "system",
+            "content": _memory_prompt_line(memory_prompt, "query_system"),
+        },
         {"role": "user", "content": user},
     ]
     try:
@@ -269,6 +284,7 @@ def analyze_for_save(
     existing_scopes: list[str],
     existing_categories: list[str],
     llm: Any,
+    memory_prompt: MemoryPromptConfig | None = None,
 ) -> MemoryAnalysis:
     """Infer scope, categories, importance, and metadata for a single memory.
 
@@ -280,17 +296,21 @@ def analyze_for_save(
         existing_scopes: Current scope paths in the memory store.
         existing_categories: Current categories in use.
         llm: The LLM instance to use.
+        memory_prompt: Optional per-step prompt strings.
 
     Returns:
         MemoryAnalysis with suggested_scope, categories, importance, extracted_metadata.
     """
-    user = _get_prompt("save_user").format(
+    user = _memory_prompt_line(memory_prompt, "save_user").format(
         content=content,
         existing_scopes=existing_scopes or ["/"],
         existing_categories=existing_categories or [],
     )
     messages = [
-        {"role": "system", "content": _get_prompt("save_system")},
+        {
+            "role": "system",
+            "content": _memory_prompt_line(memory_prompt, "save_system"),
+        },
         {"role": "user", "content": user},
     ]
     try:
@@ -322,6 +342,7 @@ def analyze_for_consolidation(
     new_content: str,
     existing_records: list[MemoryRecord],
     llm: Any,
+    memory_prompt: MemoryPromptConfig | None = None,
 ) -> ConsolidationPlan:
     """Decide insert/update/delete for a single memory against similar existing records.
 
@@ -332,6 +353,7 @@ def analyze_for_consolidation(
         new_content: The new content to store.
         existing_records: Existing records that are semantically similar.
         llm: The LLM instance to use.
+        memory_prompt: Optional per-step prompt strings.
 
     Returns:
         ConsolidationPlan with actions per record and whether to insert the new content.
@@ -345,12 +367,15 @@ def analyze_for_consolidation(
             f"- id={r.id} | scope={r.scope} | importance={r.importance:.2f} | created={created}\n"
             f"  content: {r.content[:200]}{'...' if len(r.content) > 200 else ''}"
         )
-    user = _get_prompt("consolidation_user").format(
+    user = _memory_prompt_line(memory_prompt, "consolidation_user").format(
         new_content=new_content,
         records_summary="\n\n".join(records_lines),
     )
     messages = [
-        {"role": "system", "content": _get_prompt("consolidation_system")},
+        {
+            "role": "system",
+            "content": _memory_prompt_line(memory_prompt, "consolidation_system"),
+        },
         {"role": "user", "content": user},
     ]
     try:

lib/crewai/src/crewai/memory/encoding_flow.py

@@ -314,6 +314,7 @@ def parallel_analyze(self) -> None:
                         item.content,
                         list(item.similar_records),
                         self._llm,
+                        self._config.memory_prompt,
                     )
                 elif not fields_provided and not has_similar:
                     # Group C: field resolution only
@@ -324,6 +325,7 @@ def parallel_analyze(self) -> None:
                         existing_scopes,
                         existing_categories,
                         self._llm,
+                        self._config.memory_prompt,
                     )
                 else:
                     # Group D: both in parallel
@@ -334,13 +336,15 @@ def parallel_analyze(self) -> None:
                         existing_scopes,
                         existing_categories,
                         self._llm,
+                        self._config.memory_prompt,
                     )
                     consol_futures[i] = pool.submit(
                         contextvars.copy_context().run,
                         analyze_for_consolidation,
                         item.content,
                         list(item.similar_records),
                         self._llm,
+                        self._config.memory_prompt,
                     )
 
             # Collect field-resolution results

lib/crewai/src/crewai/memory/recall_flow.py

@@ -227,6 +227,7 @@ def analyze_query_step(self) -> QueryAnalysis:
                 available,
                 scope_info,
                 self._llm,
+                self._config.memory_prompt,
             )
             self.state.query_analysis = analysis