outhad-contextkit-sdk

Thin Python client for the Outhad ContextKit cloud — long-term memory for AI agents, with multimodal + temporal-causal recall, personalised retrieval, and tenant isolation built in.

This package ships client only. There is no local memory engine, no vector store driver, no LLM provider. Calls go to the Outhad ContextKit cloud where the heavy lifting happens — billed against the API key you mint at https://contextkit.outhad.com.

Install

pip install outhad-contextkit-sdk

Quickstart

from outhad_contextkit_sdk import MemoryClient

client = MemoryClient(api_key="ock_live_...")  # or set OUTHAD_CONTEXTKIT_API_KEY

client.add(
    messages=[
        {"role": "user", "content": "Hi, my name is Alice and I love sushi."},
    ],
    user_id="alice",
)

hits = client.search(query="What food does Alice like?", user_id="alice")
for hit in hits["results"]:
    print(hit["score"], hit["memory"])

Switching the LLM at runtime

The default model is whatever the project has saved in Dashboard → Settings. You can override it on a single call without touching the project default:

client.add(
    messages=[{"role": "user", "content": "I love sushi."}],
    user_id="alice",
    llm={"provider": "anthropic", "model": "claude-haiku-4-5-20251001"},
)

# Same override shape works on search:
client.search(
    query="What food do I like?",
    user_id="alice",
    llm={"provider": "anthropic", "model": "claude-haiku-4-5-20251001"},
)

Pass api_key inside the dict to bring your own upstream key for that single call (never persisted server-side):

client.add(
    messages=[...],
    user_id="alice",
    llm={
        "provider": "anthropic",
        "model": "claude-haiku-4-5-20251001",
        "api_key": "sk-ant-…",
    },
)

Override the embedder the same way with embedder={"provider": "...", "model": "...", "api_key": "..."}.

Without overrides every call uses the project's saved defaults — set those once in the dashboard and forget about them.

Get an API key

1. Sign up at https://contextkit.outhad.com.
2. Open the dashboard → API keys.
3. Click Create key, copy the one-shot secret (looks like ock_live_<8-char-prefix>_<32-byte-secret>).
4. Either pass it as MemoryClient(api_key=…) or export it in your shell as OUTHAD_CONTEXTKIT_API_KEY.

Auth header

The SDK sends Authorization: Token <api-key> on every request. The cloud dispatcher also accepts Bearer and X-API-Key if you prefer to hit the HTTP API directly (e.g. from curl).

Host resolution

When the production default is unreachable (DNS failure / connection refused) the client transparently retries against http://localhost:8000 and pins itself there for the rest of the session — useful while you are working against make cloud-up locally. The fallback is disabled when you pass host= or set OUTHAD_CONTEXTKIT_HOST explicitly.

Async

from outhad_contextkit_sdk import AsyncMemoryClient

async def main():
    async with AsyncMemoryClient(api_key="ock_live_...") as client:
        await client.add(
            messages=[{"role": "user", "content": "I prefer dark mode."}],
            user_id="alice",
        )

API surface

• add(messages, *, user_id, agent_id, run_id, metadata, infer, llm, embedder)
• get_all(*, user_id, agent_id, run_id, limit)
• get(memory_id)
• search(query, *, user_id, agent_id, run_id, filters, limit, llm, embedder)
• update(memory_id, *, text)
• history(memory_id)
• delete(memory_id)
• delete_all(*, user_id, agent_id, run_id)

llm / embedder (optional, since 1.1.0) — per-request model override (see Switching the LLM at runtime above).

For dashboard-style operations the same client also exposes list_api_keys(), create_api_key(), revoke_api_key() — those require a Clerk session JWT and are intended for interactive tooling.

License

Apache-2.0.