LangChain described in a 21 November blog post how it uses a "knowledge factory" to keep the company knowledge its AI agents rely on up to date.
The company said it already runs such a system for its internal go-to-market and content studio agents, and that LangSmith Engine maintains one per agent. LangChain noted it is working with a few customers to bring the approach to their agents and offered assistance to those interested.
Defining knowledge versus context
LangChain distinguishes between "context" and "knowledge" in the post How to build a knowledge factory for AI agents1. Context is immediate information available during a task, while knowledge is reusable understanding maintained over time: what a metric means, which source to trust, or how to handle a particular situation.
The company argued that valuable knowledge is often tacit, implicit, and scattered across documents or employees' heads. Agents need this knowledge collected and maintained to perform tasks correctly. The post frames the concept as an extension of "software factories," where engineers build systems for agents to write code, applied now to general knowledge work.
Three categories and automated curation
LangChain outlined three categories of knowledge:
- Skills: Procedures, instructions, and scripts for accomplishing tasks in the way the company expects.
- Domain knowledge: Stable information about products, customers, competitors, and organization policies.
- Data access guidance: Instructions on which data sources to use, where data lives, and how to query it.
The post warns against manual curation, calling it a "herculean effort." Instead, LangChain proposes a system where working agents flag missing or incorrect knowledge during operation. A dedicated "curator agent" then processes these flags on a schedule, inspecting agent traces and raw sources like Notion and Slack to update the knowledge base. The company emphasized that working agents should not maintain the knowledge themselves, as doing so would cause it to become unruly.
LangChain illustrated the process with an example of an agent preparing a revenue report. If the agent notices a renamed BigQuery column, it can flag the issue and suggest the correct column name. After the finance lead works with the agent to fix formatting issues, the curator agent inspects the session traces overnight and updates the relevant skills to reflect the new column and header requirements.
Storage and standards
For storage, LangChain suggested knowledge can be kept in GitHub, Notion, Neo4j graphs, or LangSmith's Context Hub. The company noted that the industry has converged on "skills" for procedural memory, but acknowledged that no standard has yet emerged for semantic memory, adding that formats like the Open Knowledge Format (OKF) might emerge Open Knowledge Format2.
Other companies have described similar systems. Atlassian detailed an "AI work factory" that includes a feedback loop where agents write run logs back to the knowledge layer Building your AI work factory - Inside Atlassian3. Sarvam AI introduced a "knowledge fabric" for enterprise agents that uses long-running agents to review traces and improve the fabric Knowledge fabric for enterprise agents | Sarvam AI4.
The LangChain post reports no benchmarks, results, or customer outcomes. It includes no independent test results.