Python SDK

Litica Python SDK

Updated August 1, 2026·v0.1.0·Python 3.10+·GitHub

The official Python client for the Litica API. Write memories, search them, share them between agents, and inspect why a result surfaced — from code instead of an MCP client.

pip install litica

The package is litica on PyPI. It is a thin, synchronous client for the same hosted API that powers the MCP server — one method per API route, 22 routes in total. It requires Python 3.10 or newer and has a single dependency, httpx.

Two minutes to a working client

import litica

client = litica.Client(api_key="lk_...")

client.add_memory("Sam owns the Atlas pricing page.")

# Writes are queued — poll briefly before searching. See "Writing Memories".
for hit in client.search_memories("who owns pricing?"):
    print(hit.id, hit.text)

The conventional import shorthand is lit, which is why the class is Client rather than LiticaClient:

import litica as lit

client = lit.Client(api_key="lk_...")

What the SDK covers

Memories — add one, add a batch, upload a whole document, search, list, trace, delete. Writing Memories and Search & Retrieval.
Namespaces — shared memory between agents, with per-agent read and write grants. Namespaces.
Inspection — the memory graph, the write-side audit feed, queue depth, and search with the full score breakdown. Inspection.

Every method, parameter, and response type is documented in the API Reference.

Design principles

One method per route. The client mirrors the API one-to-one. Nothing is invented on top, so what you learn from the REST reference transfers directly.
Typed, tolerant responses. Responses are frozen dataclasses with the full untouched body on .raw. A server that grows a new field never breaks an older SDK.
Honest errors. Everything raised inherits from LiticaError and carries the status code and the server's message verbatim. Error Handling.
Fully typed. The package ships a py.typed marker (PEP 561), so mypy and pyright see every annotation.

Current limitations

v0.1.0 is deliberately small, and its gaps are documented rather than papered over:

Writes are queued, not instant. The server accepts a write (HTTP 202) before it is searchable, and there is no built-in wait. You poll — the pattern is in Writing Memories.
No async client and no automatic retries — out of scope for this release.
No provenance on writes. The MCP tools can record where a fact came from; the HTTP route the SDK wraps does not accept that yet.
No tenant provisioning. That route uses a separate admin credential and is intentionally absent.

While the version is 0.x the surface may still change between minor versions; breaking changes are called out explicitly in the changelog.

License and source

The SDK is open source under Apache-2.0 at Litica-AI/litica-sdk. The license covers the client library only — the Litica service is separate, and an API key is issued separately.

Next steps

Quickstart — install, get a key, write and search your first memory.
Configuration — API keys, environment variables, and how agent and namespace scope works.