Integrations / Citadel 2.2

Encrypted memory for your framework.

Framework-specific adapters for storage, sessions, and retrieval. Each requires an explicit embedder; search behavior and filters follow the adapter's interface.

Agent frameworks

Choose an adapter.

Setup examples and supported operations are documented per package. The 2.2 APIs are available in the source tree; PyPI lists published versions.

LangGraph

BaseStore

Namespace and keyed-item storage, TTL, and multi-field vector search.

Setup and API · PyPI versions

pip install citadeldb-langgraph

CrewAI

StorageBackend

Memory persistence with cosine similarity and cryptographic forgetting.

Setup and API · PyPI versions

pip install citadeldb-crewai

LangChain

VectorStore + chat history

Vector retrieval, MMR over stored vectors, metadata filters, and conversation history.

Setup and API · PyPI versions

pip install citadeldb-langchain

LlamaIndex

BasePydanticVectorStore

Encrypted vector storage with metadata filtering.

Setup and API · PyPI versions

pip install citadeldb-llamaindex

Haystack

DocumentStore

Document persistence, embedding retrieval, and metadata filters.

Setup and API · PyPI versions

pip install citadeldb-haystack

OpenAI Agents SDK

Session

Encrypted conversation history keyed by session ID.

Setup and API · PyPI versions

pip install citadeldb-openai-agents

Google ADK

BaseMemoryService

Session ingestion and hybrid memory recall.

Setup and API · PyPI versions

pip install citadeldb-google-adk

Microsoft Agent Framework

History + Context

Persistent history and memory context for agents.

Setup and API · PyPI versions

pip install citadeldb-ms-agent-framework

Strands Agents

SessionRepository

Encrypted session, agent-state, and message persistence.

Setup and API · PyPI versions

pip install citadeldb-strands-agents

Embedding and storage

Use a semantic embedder.

The default Python wheel accepts user-supplied embeddings. A custom backend provides dim, metric, model_id, and embed_with_cancel(texts, cancel_token). Check cancellation between bounded batches.

The Python guide includes a complete semantic-memory example. Framework-native embedding interfaces are supported where documented by the adapter. Candle and cross-encoder classes require a Python source build with candle-embed.

Adapters can share a database on the same thread with compatible open options. Use separate adapter-owned regions; see the shared-database guide. Cryptographic forgetting affects the active vault's keys, not external backups, replicas, or physical-media copies.

Model Context Protocol

Connect an MCP client.

The standalone server includes local Candle models and runs over stdio. Supply a vault path, passphrase, and embedder. A cross-encoder reranker is optional.

uvx citadeldb-mcp pull e5-large
uvx citadeldb-mcp pull ms-marco-minilm

Model pulls do not need a vault key. Before serving, set CITADEL_KEY to the vault passphrase: export CITADEL_KEY="your-passphrase" on macOS/Linux, or $env:CITADEL_KEY = "your-passphrase" in PowerShell.

uvx citadeldb-mcp --db memory.cdl --embedder e5-large --reranker ms-marco-minilm

For a desktop MCP client, use an absolute database path and set CITADEL_KEY in its server environment. See the client configuration. Managed downloads are pinned and verified; user-supplied model directories are outside managed-cache validation.