Building Persistent AI Memory for API Specifications Using Cognee and Fern

Description and use

A persistent knowledge graph for API specifications was built using Cognee and Fern, enabling AI agents to perform multi-hop reasoning over API architecture.

Problem

Standard vector retrieval (RAG) treats API endpoint schemas as plain text, fragmenting path parameters from query parameters, request bodies from response types, and error definitions from endpoint paths, preventing AI agents from accessing structured API knowledge.

Outcome

The Fern connector enabled AI agents to persistently remember and reason over entire API ecosystems, providing structured, multi-hop architectural memory.

Workflow

• 1. Install cognee-community-connector-fern and cognee via pip.
• 2. Point fern_source to the Fern workspace directory.
• 3. Ingest API definitions and guides with cognee.add().
• 4. Cognify the ingested data into a knowledge graph with cognee.cognify().
• 5. Query the graph using cognee.search() with INSIGHTS query type.

Evidence

Standard vector retrieval (RAG) fails. It treats endpoint schemas as plain text, fragmenting path parameters from query parameters, request bodies from response types, and error definitions from endpoint paths.

Source

We built the official Fern API Documentation Connector for Cognee.

Source

pip install cognee-community-connector-fern cognee

Source

await cognee.search(query_type="INSIGHTS", query_text="What endpoints handle user authentication and what error types can they throw?")

Source