ai-memory

name	ai-memory
description	AI Memory is a feature that allows you to store and retrieve information from your AI assistant. This can be useful for keeping track of important information, such as your schedule, contacts, and other data.
license	Complete terms in LICENSE.txt

AI Memory Guide

Overview

AI Memory is a feature that allows you to store and retrieve information from your AI assistant. This can be useful for keeping track of important information, such as your schedule, contacts, and other data.

Quickstart

Ingest your preference data into the AI Memory.

import cognee

PREFERENCE_DATA = "" # get from the app 
await cognee.add(PREFERENCE_DATA)
await cognee.cognify()

Retrieve your preference data from the AI Memory.

import asyncio
import cognee
async def main():
    user_query=""
    answers = await cognee.search(
        query_text=user_query,
        query_type=cognee.SearchType.GRAPH_COMPLETION
    )
    for answer in answers:
        print(answer)

asyncio.run(main())

output format for the retriever:

{'search_result': ['Audi was founded in the early 1900s.'], 'dataset_id': UUID('de430473-b286-5641-9081-655751d6684e'), 'dataset_name': 'main_dataset', 'dataset_tenant_id': None}

Important-Concepts

NodeSets

Tag and organize data at the dataset level. Pass tags when adding data:

await cognee.add("text", node_set=["projectA", "finance"])

What happens: Tags become graph nodes with belongs_to_set edges, enabling filtered searches within tagged subsets.

When to use:

• Group data by project, domain, or user
• Filter searches to specific subsets
• Organize memory at the dataset level

Flow:

• Add: Tags attached to datasets/documents
• Cognify: Tags materialize as NodeSet nodes in graph, entities inherit parent NodeSet links
• Search: Scope queries to specific NodeSets for targeted retrieval

Example use case: Tag patient medications as ["medications", "daily"] and emergency contacts as ["emergency"] to enable targeted queries.

Ontologies

Optional RDF/OWL file for entity validation and enrichment:

await cognee.cognify(datasets=["my_dataset"], ontology_file_path="subset.owl")

What it does: Validates and enriches extracted entities against a reference vocabulary. Matched entities are marked ontology_valid=True and inherit parent classes and relationships.

When to use:

• Standardize entity representation (e.g., "aspirin" → "medication")
• Add domain relationships from existing schemas
• Validate entities against medical/enterprise vocabularies

Format: Any RDFLib-parsable format (.owl, .ttl, .rdf, JSON-LD)

Best practice: Use small, curated subsets (50-500 classes), not full DBpedia/Wikidata. Create domain-specific ontologies or extract relevant subsets from public sources.

Sessions & Caching

Maintains conversational context across searches using (user_id, session_id):

answers = await cognee.search(
    query_text=query,
    query_type=cognee.SearchType.GRAPH_COMPLETION,
    session_id="user_123"
)

What it does: Stores short-term conversation history (Q&A pairs, context) scoped to (user_id, session_id). LLM receives both current query and previous interactions for context-aware responses.

When to use:

• Enable follow-up questions ("What about her other medications?")
• Maintain conversation context within a user session
• Support multi-turn caregiving chats

Search types that use sessions: GRAPH_COMPLETION, RAG_COMPLETION, TRIPLET_COMPLETION (batch queries do not)

Setup:

# .env
CACHING=true
CACHE_BACKEND=redis  # or filesystem
CACHE_HOST=localhost
CACHE_PORT=6379

Note: Without caching, searches work but forget previous context. If session_id is omitted, defaults to default_session.

Reference

Check out the cognee repository for more information.