langchain-goodmem

Use GoodMem from LangChain for document ingestion, retrieval, and agent memory. GoodMem handles storage, chunking, embeddings, and optional reranking.

Version 0.2 intentionally breaks compatibility with 0.1. See CHANGELOG.md for migration details.

Install and connect

Requires Python 3.10+, a running GoodMem server, an API key, and a space configured with an embedder.

pip install langchain-goodmem
export GOODMEM_BASE_URL="http://localhost:8080"
export GOODMEM_API_KEY="your-key"

Retrievers and tools read these environment variables. You can also pass a configured goodmem.Goodmem instance as client=; you retain ownership of it.

Write Documents and retrieve

import os
from goodmem import Goodmem
from langchain_core.documents import Document
from langchain_goodmem import GoodMemRetriever, add_documents

space_id = "your-space-uuid"
with Goodmem(
    base_url=os.environ["GOODMEM_BASE_URL"],
    api_key=os.environ["GOODMEM_API_KEY"],
) as client:
    memory_ids = add_documents(client, space_id, [
        Document(
            page_content="Project Cobalt's launch owner is Ada.",
            metadata={"source": "https://example.org/cobalt", "team": "blue"},
        )
    ])

retriever = GoodMemRetriever(
    space_ids=[space_id], k=5, filter="CAST(val('$.team') AS TEXT) = 'blue'",
)
for document in retriever.invoke("Who owns Cobalt's launch?"):
    print(document.page_content, document.metadata["source"])

add_documents batches text and metadata through the SDK and waits for indexing by default. Set wait=False for background ingestion, then use wait_for_memory(client, memory_id) when readiness matters. Optional Document IDs must be UUIDs; existing IDs produce conflicts. GoodMemIngestionError.created_memory_ids identifies successful writes if part of ingestion fails.

Filters use GoodMem expressions and execute on the server before retrieval. Omit filter to search all memories in the configured spaces. Searches run once; empty results return immediately.

The retriever returns Documents with source metadata, memory/chunk/space IDs, and scores. It supports LCEL, callbacks, batching, per-call k, and ainvoke through LangChain's thread executor.

If the server reports a real problem during a search — a reranker was unavailable, one space was unreachable — the Documents it did return are still returned, and each carries metadata["goodmem_partial"] = True and metadata["goodmem_statuses"] saying why. A problem that left no Documents at all returns an empty list and emits a UserWarning and a warning log line with the statuses, so it is distinguishable from a search that matched nothing. Neither case raises. Notices that carry no loss (FEATURE_DISABLED, LLM_CAPABILITY_INFERRED) are dropped; a status code this version does not know is reported as UNKNOWN.

For reranking, add reranker_id="your-reranker-uuid" and optionally fetch_k=20. Reranking requires no LLM.

Give an agent a search tool

from langchain_core.tools import create_retriever_tool

tool = create_retriever_tool(
    retriever, "search_project_records", "Search project records.",
    response_format="content_and_artifact",
)

The agent supplies only a query. Spaces and filters remain configured by the developer. Retrieved Documents are available in ToolMessage.artifact for citations.

Other tools and development

The package also provides space/memory management tools. GoodMemRetrieveMemories returns SDK events, including chunks, optional summaries, and statuses; GoodMemRetriever flags incomplete retrieval on the Documents' metadata rather than raising. Tools use SDK data shapes and LangChain ToolException handling.

See the smoke example for a complete workflow.

uv sync --all-groups
uv run pytest --disable-socket --allow-unix-socket tests/unit_tests
uv run ruff check .
uv run mypy .