Cognee — Memory Engine for AI Agents

What it is

Cognee is an open-source memory and knowledge management engine built for AI agents. While most AI tools lose context between sessions, Cognee provides persistent structured memory that grows over time. It processes text, documents, URLs, and databases into knowledge graphs with entity extraction, relationship mapping, and vector embeddings.

Cognee is designed for developers building AI agents that need to remember previous interactions, accumulate domain knowledge, or reason over structured data. It works with any LLM provider and integrates in as few as six lines of Python.

How it saves time or tokens

Without a memory engine, AI agents must re-ingest context every session, consuming tokens and adding latency. Cognee processes data once into a persistent knowledge graph, so subsequent queries retrieve only relevant facts rather than re-processing entire documents. This reduces token consumption for repeated queries and enables agents to build cumulative understanding across sessions.

The structured approach also means agents get sourced answers rather than hallucinated ones -- Cognee traces every answer back to the original data that informed it.

How to use

1. Install Cognee: pip install cognee.
2. Add data with await cognee.add('your text or document path') -- Cognee accepts text strings, file paths, and URLs.
3. Process data into the knowledge graph with await cognee.cognify(), then query with results = await cognee.search('your question').

Example

import cognee

# Add knowledge from different sources
await cognee.add('The quarterly revenue was $4.2M, up 23% YoY.')
await cognee.add('Customer churn decreased to 2.1% in Q3.')

# Process into knowledge graph
await cognee.cognify()

# Query with natural language
results = await cognee.search('What was the revenue growth?')
# Returns structured, sourced answers

Related on TokRepo

• AI memory tools -- memory solutions for AI agents compared
• AI agent tools -- frameworks for building capable AI agents

Common pitfalls

• Cognee requires an async runtime; ensure your application uses asyncio or an async-compatible framework when calling cognee.add() and cognee.cognify().
• Large document batches should be processed incrementally rather than all at once to avoid memory pressure during the knowledge graph construction phase.
• The default vector store works for development but should be replaced with a production-grade store (PostgreSQL with pgvector, Weaviate, or similar) for deployment.