zeno-chroma

ChromaDB-backed KnowledgeStore adapter for Zeno.

Persists facts and documents in a local Chroma collection with each row stamped by user_id so multi-tenant isolation is enforced at query time via Chroma's where filter rather than post-fetch filtering. Plugs into the ThreeLayer memory binder as the knowledge and (optionally) the user_memory backing.

Install

uv add 'zeno-framework[chroma]'
# or, without the AI package:
uv add zeno-chroma

Features

• Persistent local store via chromadb.PersistentClient — no server to run, just a directory
• user_id-scoped search using Chroma's where filter (S7 isolation is structural, not advisory)
• Cosine similarity, distance → similarity score conversion (1.0 - distance, clamped to [0, 1])
• Synchronous Chroma client wrapped in asyncio.to_thread so embeddings and queries don't block the event loop
• Default Chroma embedding function; pass embedding_function= to override
• Pairs with VectorBackedUserMemoryStore to also serve user-memory reads/writes from the same collection

Minimal usage

import asyncio
from pathlib import Path

from zeno.agent import Agent
from zeno.app import ZenoApp
from zeno.builtin_tools import recall, remember
from zeno.channels.cli import CliChannel
from zeno.chroma import ChromaKnowledgeStore
from zeno.memory.sqlite.conversation_store import SqliteConversationStore
from zeno.memory.sqlite.session_store import SqliteSessionStore
from zeno.memory.three_layer import ThreeLayer
from zeno.memory.vector_backed.user_memory_store import VectorBackedUserMemoryStore
from zeno.testing import FakeProvider

async def main() -> None:
    data_dir = Path("./.zeno")
    data_dir.mkdir(exist_ok=True)

    knowledge = ChromaKnowledgeStore(persist_dir=data_dir / "chroma")
    memory = ThreeLayer(
        session=SqliteSessionStore(data_dir / "sessions.db"),
        user_memory=VectorBackedUserMemoryStore(knowledge_store=knowledge),
        knowledge=knowledge,
        conversation=SqliteConversationStore(data_dir / "conversations.db"),
    )

    app = ZenoApp(
        agent=Agent(
            name="root",
            instructions="Use `recall` to look things up and `remember` to store facts.",
            tools=[recall, remember],
        ),
        channels=[CliChannel()],
        provider=FakeProvider(),
        memory=memory,
    )
    await app.run()

asyncio.run(main())

The first call to remember writes a row tagged with the active user_id. Subsequent recall calls only see rows that user inserted — cross-user isolation is enforced by the where={"user_id": ...} filter on every query.

Testing

uv run pytest packages/zeno-chroma

Tests run against a temporary Chroma directory so no global state leaks between cases. The contract suite shared with zeno-qdrant guarantees both adapters honour the same KnowledgeStore semantics.

See also

• zeno-qdrant — the other first-party KnowledgeStore adapter.
• docs/memory.md — three-layer memory model.
• docs/adapters.md — writing your own KnowledgeStore.

Part of the Zeno framework.