citadeldb-openai-agents

An OpenAI Agents SDK Session backed by Citadel. Encrypted at rest, embedded in your process, and deletes that destroy the key, not just the row.

pip install citadeldb-openai-agents

from agents import Agent, Runner
from citadeldb_openai_agents import CitadelSession

session = CitadelSession("user-123", "agent.cdl", key="your-passphrase")

agent = Agent(name="assistant", instructions="Be brief.")
result = await Runner.run(agent, "remember my dog is called Mochi", session=session)
result = await Runner.run(agent, "what is my dog called?", session=session)

The conversation persists across processes, so the second run answers from the transcript rather than from the prompt.

Many sessions, one file

Citadel is embedded and one connection owns the file, so sessions are minted from a store rather than each opening the database:

from citadeldb_openai_agents import CitadelSessionStore

store = CitadelSessionStore("agent.cdl", key="your-passphrase")
alice = store.session("user-alice")
bob = store.session("user-bob")

The convenience constructor does this for you: two CitadelSession objects on the same path build their own store but reach one open database. Asking for the same path with a different passphrase raises rather than quietly serving the first one's settings.

Deletes destroy the key

Every item is sealed under its own key. clear_session destroys those keys, so the bytes on disk stay unreadable. A backup taken before the delete carries its own copy of the wrapped key and is out of scope.

await session.clear_session()

pop_item does the same for a single rolled-back turn.

The SDK's own EncryptedSession wrapper encrypts items and skips expired ones on read, but the ciphertext and its key both remain.

Search the transcript

Beyond the protocol, a session can be searched with Citadel's hybrid recall, which ranks on vector distance, keyword rank and recency rather than on an exact match:

await session.add_items([
    {"role": "user", "content": "the deployment failed because the disk was full"},
    {"role": "user", "content": "lunch plans for friday"},
])

await session.search("why did the release break?", limit=1)
# [{'content': 'the deployment failed because the disk was full', 'role': 'user'}]

Nothing in the SDK calls this. Runner only ever uses the four protocol methods.

TTL

store = CitadelSessionStore("agent.cdl", key="your-passphrase", ttl=86400)  # seconds

Expired items stop being returned and are skipped by the storage engine itself, so a retention window needs no sweeper.

Notes

Items are stored verbatim as opaque JSON. The SDK's item type is a large union owned by the openai package, so the stored payload is never normalised: function calls, reasoning items and multi-part content all round-trip unchanged. Only a plain-text projection of content is derived, for search ranking.

MockEmbedder is the default and needs no download, which is enough to run an agent and to test. search only becomes semantically useful with a real embedder. CandleEmbedder is not in the default citadeldb wheel and needs a source build (maturin build --features candle-embed); any object exposing dim, metric, model_id, embed and embed_queries works too:

import citadeldb
store = CitadelSessionStore(
    "agent.cdl",
    key="your-passphrase",
    embedder=citadeldb.CandleEmbedder("/path/to/e5-large", preset="e5-large"),
)

License

Apache-2.0