Where we stand: the package works and passes LangChain's standard tests, and its feature set is the December 2024 build-out — vector store, chat history, two caches, a loader. Since then the repo has been maintained, with dependency and compatibility fixes, while nearly all feature work stayed where it started. The released 0.2.0 (December 2025) pins langchain <1.2, so it cannot be installed next to current LangChain; the fixes sit on main, unreleased. Adoption is 141 installs a month.
Where the ecosystem went: LangChain 1.0 (October 2025) rebuilt the framework around agents running on LangGraph, with memory being the new valuable thing a database provides an agent with. For memory, two standard shapes: a checkpointer, which saves the agent's working state so a conversation can pause, resume and survive crashes, and a store, which holds long-term memories across conversations and searches them semantically. Postgres, SQLite, MongoDB and Redis have official implementations, the Postgres checkpointer alone pulls 8.5M installs a month, and our chat-history component plugs into a pattern that is formally deprecated. Meanwhile the class our vector store inherits from, PGVector, was deprecated by langchain-postgres itself in favor of a v2 with first-class async, hybrid search and index management.
The unsold strength: CrateDB holds vectors, BM25 full-text, structured metadata and time-series in one engine, joinable in one SQL query, with hybrid search documented in its own guide. Through LangChain none of that is visible — the integration exposes plain vector search, making CrateDB look like a narrower pgvector exactly where AI builders meet it. We are also absent from langchain's vector-store comparison table.
The directions, largest first: agent memory (a LangGraph checkpointer and store), a modern vector store (async, off the deprecated base), search (which crate core can already do (full-text and hybrid, #10)), presence where users choose (comparison table, provider page, a SQL-agent guide), and shipping what already exists (release repair). Then a self-query translator (an LLM turning "articles about X from last month" into metadata filters — a natural fit for a time-series database), and a key-value docstore plus record manager for parent-document retrieval. Three items need conversations outside this repo: the 2048-dimension vector cap that breaks current OpenAI and Gemini embedding defaults (#14), euclidean-only vector similarity (crate/crate#15768), and upsert support that belongs upstream in the dialect (sqlalchemy-cratedb#186).
The numbers (collected September 2, 2026)
| package |
role |
installs, 30 days |
| langgraph-checkpoint-postgres |
agent memory on Postgres |
8,495,263 |
| langchain-postgres |
Postgres vector store |
1,527,142 |
| langgraph-checkpoint-redis |
agent memory on Redis |
1,027,748 |
| langchain-mongodb |
MongoDB integration |
888,359 |
| langgraph-checkpoint-mongodb |
agent memory on MongoDB |
409,656 |
| timescale-vector |
Timescale vector client |
2,917 |
| langchain-singlestore |
SingleStore integration |
621 |
| langchain-cratedb |
CrateDB integration |
141 |
Two readings: demand concentrated in agent memory (the Postgres checkpointer outdraws the Postgres vector store five to one), and every niche vendor package is small, SingleStore's included — volume flows through ecosystem defaults and documentation tables, so presence on those pages is its own work item.
What the peers prove is possible
langchain-postgres rewrote its store as a v2 (one async implementation with a sync facade, HNSW/IVFFlat index management, hybrid search with rank fusion, migration tooling) and deprecated the class we build on. MongoDB ships checkpointer and store marketed as agent memory, five retrievers including hybrid and GraphRAG, and an agent toolkit. Redis sells the same persistence pair. Timescale markets vector search beside time-series data — our claim, their page. SingleStore leads its LangChain docs with weighted hybrid search our engine could match. CockroachDB proved the LangGraph checkpointer port for a Postgres-compatible-but-not-Postgres engine — the template for ours.
Where we stand: the package works and passes LangChain's standard tests, and its feature set is the December 2024 build-out — vector store, chat history, two caches, a loader. Since then the repo has been maintained, with dependency and compatibility fixes, while nearly all feature work stayed where it started. The released 0.2.0 (December 2025) pins
langchain <1.2, so it cannot be installed next to current LangChain; the fixes sit onmain, unreleased. Adoption is 141 installs a month.Where the ecosystem went: LangChain 1.0 (October 2025) rebuilt the framework around agents running on LangGraph, with memory being the new valuable thing a database provides an agent with. For memory, two standard shapes: a checkpointer, which saves the agent's working state so a conversation can pause, resume and survive crashes, and a store, which holds long-term memories across conversations and searches them semantically. Postgres, SQLite, MongoDB and Redis have official implementations, the Postgres checkpointer alone pulls 8.5M installs a month, and our chat-history component plugs into a pattern that is formally deprecated. Meanwhile the class our vector store inherits from,
PGVector, was deprecated by langchain-postgres itself in favor of a v2 with first-class async, hybrid search and index management.The unsold strength: CrateDB holds vectors, BM25 full-text, structured metadata and time-series in one engine, joinable in one SQL query, with hybrid search documented in its own guide. Through LangChain none of that is visible — the integration exposes plain vector search, making CrateDB look like a narrower pgvector exactly where AI builders meet it. We are also absent from langchain's vector-store comparison table.
The directions, largest first: agent memory (a LangGraph checkpointer and store), a modern vector store (async, off the deprecated base), search (which crate core can already do (full-text and hybrid, #10)), presence where users choose (comparison table, provider page, a SQL-agent guide), and shipping what already exists (release repair). Then a self-query translator (an LLM turning "articles about X from last month" into metadata filters — a natural fit for a time-series database), and a key-value docstore plus record manager for parent-document retrieval. Three items need conversations outside this repo: the 2048-dimension vector cap that breaks current OpenAI and Gemini embedding defaults (#14), euclidean-only vector similarity (crate/crate#15768), and upsert support that belongs upstream in the dialect (sqlalchemy-cratedb#186).
The numbers (collected September 2, 2026)
Two readings: demand concentrated in agent memory (the Postgres checkpointer outdraws the Postgres vector store five to one), and every niche vendor package is small, SingleStore's included — volume flows through ecosystem defaults and documentation tables, so presence on those pages is its own work item.
What the peers prove is possible
langchain-postgres rewrote its store as a v2 (one async implementation with a sync facade, HNSW/IVFFlat index management, hybrid search with rank fusion, migration tooling) and deprecated the class we build on. MongoDB ships checkpointer and store marketed as agent memory, five retrievers including hybrid and GraphRAG, and an agent toolkit. Redis sells the same persistence pair. Timescale markets vector search beside time-series data — our claim, their page. SingleStore leads its LangChain docs with weighted hybrid search our engine could match. CockroachDB proved the LangGraph checkpointer port for a Postgres-compatible-but-not-Postgres engine — the template for ours.