Cognee

Overview

Cognee solves the memory problem that plagues every AI agent: context windows are finite and stateless. Feed it documents, text, or any data format, and it builds a dual-index — knowledge graphs for relationships, vector stores for semantic search — that agents can query for relevant context on demand.

The pitch: 6 lines of code to give your agent persistent memory. The reality: those 6 lines hide a serious pipeline underneath — Extract, Cognify, Load — that handles chunking, entity extraction, relationship mapping, and indexing.

flowchart LR
    A[Raw Data] --> B[Add]
    B --> C[Cognify]
    C --> D[Knowledge Graph]
    C --> E[Vector Store]
    D --> F[Search]
    E --> F
    F --> G[AI Agent Context]

    subgraph ECL Pipeline
        B
        C
    end

    subgraph Storage
        D
        E
    end

Key Features

• ECL Pipeline — Extract, Cognify, Load. Data goes in raw, comes out as structured knowledge with entities, relationships, and embeddings
• Dual search — Combines graph-based traversal (follow relationships) with vector similarity (find meaning). Neither alone is sufficient
• Multi-tenant isolation — User and tenant separation baked in, not bolted on. Critical for production agent systems
• Ontology grounding — Knowledge isn't just extracted, it's mapped to defined concepts. Reduces hallucination by anchoring to structure
• Observability — OTEL collector integration and audit trails. You can trace why an agent got a specific piece of context

Code Snippets

Installation

pip install cognee

Core API — 6 Lines of Memory

The entire add-cognify-search loop in one shot.

import cognee
import asyncio

async def main():
    await cognee.add("Cognee turns documents into AI memory.")
    await cognee.cognify()
    results = await cognee.search("What does Cognee do?")
    for result in results:
        print(result)

asyncio.run(main())

CLI Alternative

cognee-cli add "Cognee turns documents into AI memory."
cognee-cli cognify
cognee-cli search "What does Cognee do?"

Technical Details

Built on Python (93.4%), async-first architecture. Uses Neo4j for graph storage by default, with vector database support for embedding-based retrieval. Supports multiple LLM providers — configure via environment variables or .env file. Requires Python 3.10-3.13.

The architecture leans on cognitive science concepts — not just vector similarity, but structured knowledge representation. The "cognify" step does the heavy lifting: chunking, entity recognition, relationship extraction, and embedding generation in a single pipeline pass.

Connections

• advanced-context-engineering-for-coding-agents - Both tackle the same core problem from different angles: how do you give AI agents the right context? That repo focuses on prompt-level context management, Cognee attacks it at the infrastructure layer with persistent knowledge graphs
• ai-assisted-development - Cognee is the kind of infrastructure that makes AI-assisted development actually work at scale — agents that remember previous sessions, understand your codebase relationships, and don't lose context mid-conversation