# LangChain describes a knowledge factory for keeping AI agents current

Published 2026-11-21 · AI.info Team · AI.info

> LangChain says it uses a system in which agents flag knowledge gaps and a separate agent updates shared information. The company is also working with a few customers to bring the approach to their agents.

[LangChain](https://www.langchain.com/blog/knowledge-factory) says its internal go-to-market and content studio agents run on what it calls a knowledge factory: a system to collect, curate and maintain the company knowledge its agents need. In a November 21 post, Harrison Chase described how the company uses it for internal work and in LangSmith Engine.

The company is also working with a few customers to bring the approach to their internal and external agents, Chase wrote. LangSmith Engine, an agent that examines traces and suggests fixes, maintains a knowledge factory for each agent, according to the post.

## Knowledge for future tasks

Chase distinguishes the context available to an agent during a task from reusable knowledge about how a company operates. That knowledge can include the meaning of a metric, which source to trust or how to handle a particular situation.

A knowledge factory is intended to collect and maintain that information so it can be supplied as context for later work. Other context that is already reliably stored can be provided directly to an agent, rather than maintained separately in the factory.

Chase divides that knowledge into skills, domain knowledge and guidance on accessing data. Skills cover procedures, instructions and scripts for doing work in the way a company expects.

Domain knowledge covers definitions, policies and facts about products, customers, competitors and the organization. The third category tells agents which sources to use, where information lives and how to query it.

## A curator in the background

Agents doing their usual work would flag information they cannot find or discover is wrong, and could suggest a correction. A separate curator agent would periodically review those flags and update the knowledge store.

Chase argues that the agents using the knowledge should not be solely responsible for maintaining it: asking agents that are mainly pursuing their own goals to maintain it “will cause it to become unruly and full of contradictions,” he wrote. The curator would instead have access to the underlying sources, as well as traces showing how people use agents.

Those sources could include Notion and Slack. Traces could help the curator turn interactions into skills, while reports from working agents would identify material that needs attention.

Possible storage formats include skills in [GitHub](https://ai.info/companies/github), a file in Notion, OKF format in LangSmith Context Hub or a knowledge graph in Neo4j. Chase wrote that the representation of semantic memory remains an early area without an established standard.

## A revenue-report example

His post illustrates the process with an agent preparing a revenue report. In the example, the raw revenue figures sit in BigQuery and [Salesforce](https://ai.info/companies/salesforce), while reporting instructions are stored as skills in GitHub and the company's method for calculating ARR is kept in a wiki.

Instructions for accessing the changing data are stored as skills in Context Hub. In that example, the reporting agent notices a renamed BigQuery column, finds the correct one while working and flags the issue with a suggested update.

After the report is created, the finance lead works with the agent on formatting, including a new header for company documents. The curator then reviews the flagged column and the agent's trace that night, updating the relevant skills for both the data query and the report header.

## Sources

1. Original source · Blog post [LangChain](https://www.langchain.com/blog/knowledge-factory) langchain.com

## Companies in this story

- [LangChain Inc.](https://ai.info/companies/langchain-inc)

## Related stories

- [AWS Quick adds multi-agent responses and chat-built business flows](https://ai.info/news/aws-quick-multi-agent-business-flows)
- [Cohere brings Crunchbase data to North for licensed customers](https://ai.info/news/cohere-north-crunchbase-connector)
- [SOCi makes MCP server available to all customers](https://ai.info/news/soci-mcp-server-integrations-agentic-workforce)
- [35Gates offers dedicated cloud AI environments from ¥64,000 a month before tax](https://ai.info/news/35gates-lmaas-dedicated-cloud-ai-platform)
- [Architect Labs says its AI systems designed the Frostbite CPU](https://ai.info/news/architect-labs-frostbite-ai-designed-cpu)

## Explore

[More articles](https://ai.info/news)

---

Canonical: https://ai.info/news/langchain-knowledge-factory-ai-agents
