lians

Lians Memory

Lians is a memory layer for AI agents built for regulated environments. Unlike a plain vector store, it uses a bitemporal model: every fact carries the business time it became true (event_time) and the system time it was known (valid_from/valid_to). When a new fact supersedes an old one, the old one is excluded from recall automatically — but remains reconstructable for any past date.

Use Lians when the agent works with facts that change over time: guidance revisions, dosage changes, matter status, prior decisions — and when those changes must be auditable.

Setup

pip install lians-sdk          # hosted / self-hosted
pip install lians-sdk[local]   # zero-setup local SQLite (no server, no API key)

Environment (hosted/self-hosted): LIANS_URL, LIANS_API_KEY, LIANS_AGENT_ID. Get a free key at api.lians.dev. Local mode needs none.

Clients (same API surface)

from lians import LiansClient        # sync HTTP — scripts, CLIs
from lians import AsyncLiansClient   # async HTTP — FastAPI, async frameworks
from lians import LocalLiansClient   # local SQLite — prototyping, CI, notebooks

Core operations

from datetime import datetime, timezone
mem = LiansClient(base_url=os.environ["LIANS_URL"], api_key=os.environ["LIANS_API_KEY"])

# Store — event_time is the BUSINESS time the fact became true, not now
mem.add(agent_id="desk", content="NVDA guidance raised to $40B",
        event_time=datetime(2025, 11, 19, tzinfo=timezone.utc),
        metadata={"ticker": "NVDA", "metric": "revenue_guidance"})

# Recall — current, non-stale facts only
res = mem.recall(agent_id="desk", query="NVDA guidance", k=5)
for m in res["memories"]:
    print(m["event_time"], m["content"])

# Point-in-time — what did we know on a past date?
mem.recall_at(agent_id="desk", query="NVDA guidance",
              as_of=datetime(2025, 9, 1, tzinfo=timezone.utc))

# From a conversation
mem.add_from_messages(agent_id="desk",
    messages=[{"role": "assistant", "content": "TSLA Q4 deliveries hit 495k"}])