Add a long-term memory interface

CLAUDE.md

@@ -42,6 +42,7 @@ docker-compose down          # Stop all services
 IMPORTANT: This project uses `pre-commit`. You should run `pre-commit`
 before committing:
 ```bash
+uv run pre-commit install  # Install the hooks first
 uv run pre-commit run --all-files
 ```
 
@@ -68,7 +69,7 @@ Working Memory (Session-scoped)  →  Long-term Memory (Persistent)
 ```python
 # Correct - Use RedisVL queries
 from redisvl.query import VectorQuery, FilterQuery
-query = VectorQuery(vector=embedding, vector_field_name="embedding", return_fields=["text"])
+query = VectorQuery(vector=embedding, vector_field_name="vector", return_fields=["text"])
 
 # Avoid - Direct redis client searches
 # redis.ft().search(...)  # Don't do this

README.md

@@ -15,6 +15,16 @@ A Redis-powered memory server built for AI agents and applications. It manages b
 - **Long-Term Memory**
 
   - Persistent storage for memories across sessions
+  - **Pluggable Vector Store Backends** - Support for multiple vector databases through LangChain VectorStore interface:
+    - **Redis** (default) - RedisStack with RediSearch
+    - **Chroma** - Open-source vector database
+    - **Pinecone** - Managed vector database service
+    - **Weaviate** - Open-source vector search engine
+    - **Qdrant** - Vector similarity search engine
+    - **Milvus** - Cloud-native vector database
+    - **PostgreSQL/PGVector** - PostgreSQL with vector extensions
+    - **LanceDB** - Embedded vector database
+    - **OpenSearch** - Open-source search and analytics suite
   - Semantic search to retrieve memories with advanced filtering system
   - Filter by session, namespace, topics, entities, timestamps, and more
   - Supports both exact match and semantic similarity search
@@ -84,6 +94,8 @@ Configure servers and workers using environment variables. Includes background t
 
 For complete configuration details, see [Configuration Guide](docs/configuration.md).
 
+For vector store backend options and setup, see [Vector Store Backends](docs/vector-store-backends.md).
+
 ## License
 
 Apache 2.0 License - see [LICENSE](LICENSE) file for details.

agent-memory-client/agent_memory_client/client.py

@@ -13,8 +13,8 @@
     from typing_extensions import Self
 
 import httpx
-import ulid
 from pydantic import BaseModel
+from ulid import ULID
 
 from .exceptions import MemoryClientError, MemoryServerError, MemoryValidationError
 from .filters import (
@@ -466,7 +466,7 @@ async def add_memories_to_working_memory(
         # Auto-generate IDs for memories that don't have them
         for memory in final_memories:
             if not memory.id:
-                memory.id = str(ulid.ULID())
+                memory.id = str(ULID())
 
         # Create new working memory with the memories
         working_memory = WorkingMemory(

agent-memory-client/agent_memory_client/models.py

@@ -9,8 +9,8 @@
 from enum import Enum
 from typing import Any, Literal, TypedDict
 
-import ulid
 from pydantic import BaseModel, Field
+from ulid import ULID
 
 # Model name literals for model-specific window sizes
 ModelNameLiteral = Literal[
@@ -122,7 +122,7 @@ class ClientMemoryRecord(MemoryRecord):
     """A memory record with a client-provided ID"""
 
     id: str = Field(
-        default_factory=lambda: str(ulid.ULID()),
+        default_factory=lambda: str(ULID()),
         description="Client-provided ID generated by the client (ULID)",
     )
 

agent_memory_server/api.py

@@ -1,8 +1,8 @@
 import tiktoken
-import ulid
 from fastapi import APIRouter, Depends, HTTPException
 from mcp.server.fastmcp.prompts import base
 from mcp.types import TextContent
+from ulid import ULID
 
 from agent_memory_server import long_term_memory, working_memory
 from agent_memory_server.auth import UserInfo, get_current_user
@@ -344,7 +344,7 @@ async def put_working_memory(
 
             memories = [
                 MemoryRecord(
-                    id=str(ulid.ULID()),
+                    id=str(ULID()),
                     session_id=session_id,
                     text=f"{msg.role}: {msg.content}",
                     namespace=updated_memory.namespace,
@@ -449,13 +449,10 @@ async def search_long_term_memory(
     if not settings.long_term_memory:
         raise HTTPException(status_code=400, detail="Long-term memory is disabled")
 
-    redis = await get_redis_conn()
-
     # Extract filter objects from the payload
     filters = payload.get_filters()
 
     kwargs = {
-        "redis": redis,
         "distance_threshold": payload.distance_threshold,
         "limit": payload.limit,
         "offset": payload.offset,
@@ -465,7 +462,7 @@ async def search_long_term_memory(
     if payload.text:
         kwargs["text"] = payload.text
 
-    # Pass text, redis, and filter objects to the search function
+    # Pass text and filter objects to the search function (no redis needed for vectorstore adapter)
     return await long_term_memory.search_long_term_memories(**kwargs)
 
 
@@ -635,6 +632,16 @@ async def memory_prompt(
                     ),
                 )
             )
+        else:
+            # Always include a system message about long-term memories, even if empty
+            _messages.append(
+                SystemMessage(
+                    content=TextContent(
+                        type="text",
+                        text="## Long term memories related to the user's query\n No relevant long-term memories found.",
+                    ),
+                )
+            )
 
     _messages.append(
         base.UserMessage(

agent_memory_server/config.py

@@ -1,5 +1,5 @@
 import os
-from typing import Literal
+from typing import Any, Literal
 
 import yaml
 from dotenv import load_dotenv
@@ -9,12 +9,42 @@
 load_dotenv()
 
 
-def load_yaml_settings():
-    config_path = os.getenv("APP_CONFIG_FILE", "config.yaml")
-    if os.path.exists(config_path):
-        with open(config_path) as f:
-            return yaml.safe_load(f) or {}
-    return {}
+# Model configuration mapping
+MODEL_CONFIGS = {
+    "gpt-4o": {"provider": "openai", "embedding_dimensions": None},
+    "gpt-4o-mini": {"provider": "openai", "embedding_dimensions": None},
+    "gpt-4": {"provider": "openai", "embedding_dimensions": None},
+    "gpt-3.5-turbo": {"provider": "openai", "embedding_dimensions": None},
+    "text-embedding-3-small": {"provider": "openai", "embedding_dimensions": 1536},
+    "text-embedding-3-large": {"provider": "openai", "embedding_dimensions": 3072},
+    "text-embedding-ada-002": {"provider": "openai", "embedding_dimensions": 1536},
+    "claude-3-opus-20240229": {"provider": "anthropic", "embedding_dimensions": None},
+    "claude-3-sonnet-20240229": {"provider": "anthropic", "embedding_dimensions": None},
+    "claude-3-haiku-20240307": {"provider": "anthropic", "embedding_dimensions": None},
+    "claude-3-5-sonnet-20240620": {
+        "provider": "anthropic",
+        "embedding_dimensions": None,
+    },
+    "claude-3-5-sonnet-20241022": {
+        "provider": "anthropic",
+        "embedding_dimensions": None,
+    },
+    "claude-3-5-haiku-20241022": {
+        "provider": "anthropic",
+        "embedding_dimensions": None,
+    },
+    "claude-3-7-sonnet-20250219": {
+        "provider": "anthropic",
+        "embedding_dimensions": None,
+    },
+    "claude-3-7-sonnet-latest": {"provider": "anthropic", "embedding_dimensions": None},
+    "claude-3-5-sonnet-latest": {"provider": "anthropic", "embedding_dimensions": None},
+    "claude-3-5-haiku-latest": {"provider": "anthropic", "embedding_dimensions": None},
+    "claude-3-opus-latest": {"provider": "anthropic", "embedding_dimensions": None},
+    "o1": {"provider": "openai", "embedding_dimensions": None},
+    "o1-mini": {"provider": "openai", "embedding_dimensions": None},
+    "o3-mini": {"provider": "openai", "embedding_dimensions": None},
+}
 
 
 class Settings(BaseSettings):
@@ -27,16 +57,28 @@ class Settings(BaseSettings):
     port: int = 8000
     mcp_port: int = 9000
 
+    # Vector store factory configuration
+    # Python dotted path to function that returns VectorStore or VectorStoreAdapter
+    # Function signature: (embeddings: Embeddings) -> Union[VectorStore, VectorStoreAdapter]
+    # Examples:
+    #   - "agent_memory_server.vectorstore_factory.create_redis_vectorstore"
+    #   - "my_module.my_vectorstore_factory"
+    #   - "my_package.adapters.create_custom_adapter"
+    vectorstore_factory: str = (
+        "agent_memory_server.vectorstore_factory.create_redis_vectorstore"
+    )
+
+    # RedisVL configuration (used by default Redis factory)
+    redisvl_index_name: str = "memory_records"
+
     # The server indexes messages in long-term memory by default. If this
     # setting is enabled, we also extract discrete memories from message text
     # and save them as separate long-term memory records.
     enable_discrete_memory_extraction: bool = True
 
     # Topic modeling
     topic_model_source: Literal["BERTopic", "LLM"] = "LLM"
-    topic_model: str = (
-        "MaartenGr/BERTopic_Wikipedia"  # Use an LLM model name here if using LLM
-    )
+    topic_model: str = "gpt-4o-mini"
     enable_topic_extraction: bool = True
     top_k_topics: int = 3
 
@@ -45,10 +87,11 @@ class Settings(BaseSettings):
     enable_ner: bool = True
 
     # RedisVL Settings
+    # TODO: Adapt to vector store settings
     redisvl_distance_metric: str = "COSINE"
     redisvl_vector_dimensions: str = "1536"
-    redisvl_index_name: str = "memory"
-    redisvl_index_prefix: str = "memory"
+    redisvl_index_prefix: str = "memory_idx"
+    redisvl_indexing_algorithm: str = "HNSW"
 
     # Docket settings
     docket_name: str = "memory-server"
@@ -74,9 +117,54 @@ class Settings(BaseSettings):
     class Config:
         env_file = ".env"
         env_file_encoding = "utf-8"
-        extra = "ignore"  # Ignore extra fields in YAML/env
+        extra = "ignore"  # Ignore extra environment variables
+
+    @property
+    def generation_model_config(self) -> dict[str, Any]:
+        """Get configuration for the generation model."""
+        return MODEL_CONFIGS.get(self.generation_model, {})
+
+    @property
+    def embedding_model_config(self) -> dict[str, Any]:
+        """Get configuration for the embedding model."""
+        return MODEL_CONFIGS.get(self.embedding_model, {})
+
+    def load_yaml_config(self, config_path: str) -> dict[str, Any]:
+        """Load configuration from YAML file."""
+        if not os.path.exists(config_path):
+            return {}
+        with open(config_path) as f:
+            return yaml.safe_load(f) or {}
+
+
+settings = Settings()
+
+
+def get_config():
+    """Get configuration from environment and settings files."""
+    config_data = {}
+
+    # If REDIS_MEMORY_CONFIG is set, load config from file
+    config_file = os.getenv("REDIS_MEMORY_CONFIG")
+    if config_file:
+        try:
+            with open(config_file) as f:
+                if config_file.endswith((".yaml", ".yml")):
+                    config_data = yaml.safe_load(f) or {}
+                else:
+                    # Assume JSON
+                    import json
+
+                    config_data = json.load(f) or {}
+        except FileNotFoundError:
+            print(f"Warning: Config file {config_file} not found")
+        except Exception as e:
+            print(f"Warning: Error loading config file {config_file}: {e}")
 
+    # Environment variables override file config
+    for key, value in os.environ.items():
+        if key.startswith("REDIS_MEMORY_"):
+            config_key = key[13:].lower()  # Remove REDIS_MEMORY_ prefix
+            config_data[config_key] = value
 
-# Load YAML config first, then let env vars override
-yaml_settings = load_yaml_settings()
-settings = Settings(**yaml_settings)
+    return config_data