How to use the Ingest data with remember() agent skill

What the cognee-ingestion skill makes the agent do

cognee is a library that stores data and builds a searchable knowledge graph from it. The remember() function is its ingestion API: one call stores the data, builds the graph and enriches it. This skill is the know-how for that function. Once it is installed, the agent can choose the right input type, route files to the correct loader, pick datasets and node_sets, decide between an LLM and a local GLiNER model for entity extraction, load ontologies, and estimate token cost with a dry run before spending money.

Install the skill

Skills are folders that an agent can read. You can install this one in either of two ways. The first is to download the ZIP of the cognee-ingestion skill from the page where you found this guide, then unzip it into ~/.claude/skills/ for your user account, or into .claude/skills/ inside a project to make it available only there. The second is to upload the ZIP in the Claude apps under Settings, then Capabilities, then Skills.

The skill's files are SKILL.md. Check that the folder contains SKILL.md after unzipping so the agent can find its instructions.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into .claude/skills/ in your project
• Claude apps: Settings, Capabilities, Skills, then upload the ZIP

Trigger the skill with natural requests

You do not need to name the skill. The skill describes when it applies: putting data into cognee memory with remember(), choosing inputs such as text, files, folders, URLs, repositories or databases, datasets and node_sets, loaders, ontologies, the graph extractor (LLM or GLiNER), chunking, dry-run cost estimates, or a remember() call that raises on a keyword argument. Ask in plain terms that name one of these tasks.

• Load this folder of PDFs and markdown notes into a dataset called research.
• Estimate the token cost of ingesting these documents before you run anything.
• Ingest these CSV files with the correct loader and tag them with the AI node set.
• Build the graph locally without an LLM using GLiNER.
• My remember() call fails with an unexpected keyword argument. What do I change?

Ingest data with remember(): basic calls

The data argument accepts a string, a list of strings, file paths (absolute, file://, or s3://), http(s) URLs, binary streams, or a list mixing them. Without dataset_name, data goes to main_dataset. All cognee functions are async, so call them with await. The skill's examples use the following pattern.

• Text: pass a string directly.
• Files: pass a list of paths and set dataset_name to keep them together.
• Folders are ingested file by file. A folder that looks like a code project, or a GitHub or GitLab URL, becomes one code repository and needs git on PATH.
• Code files such as .py or .ts go down a code-graph route with no LLM calls and are searchable only with SearchType.CODE.

Control datasets, node_sets and the graph build

A dataset is the unit of permissions and isolation, so use dataset_name or dataset_id to separate data. dataset_id wins if both are given. node_set tags data so that recall can filter to it later. Setting extractor chooses how entities are extracted: 'llm' or 'gliner_demo' (alias 'gliner'). By default, GRAPH_EXTRACTOR=auto uses the LLM when an API key is configured and GLiNER otherwise. Setting dry_run=True returns a token and cost estimate without ingesting anything or calling the LLM. It does not cover the calls that improve() makes, and it is not supported with GLiNER, sessions or a remote instance.

• dataset_name or dataset_id: target dataset
• node_set=['AI', 'FinTech']: filter later with recall(..., node_name=['AI'])
• extractor: 'llm' or 'gliner_demo'
• chunk_size and chunker: tokens per chunk and the chunker class (default TextChunker)
• dry_run=True: estimate only, no ingestion and no LLM calls

Add ontologies and choose loaders

An ontology grounds entities in an OWL file. Pass it through config with RDFLibOntologyResolver, or set ONTOLOGY_FILE_PATH, ONTOLOGY_MODE and MATCHING_STRATEGY in .env. The default mode annotate only enriches the graph. The strict mode drops entities that match no ontology class or individual, and it prunes only the graph, so chunk text is still stored.

Each file is claimed by the first loader that accepts it. You can choose a loader per file type with preferred_loaders. Office formats such as DOCX and PPTX need the docs (unstructured) or docling extra. A preferred loader that is not installed is skipped with only an info log, so check the extra when a file comes out wrong.

• ontology_mode 'annotate' (default): enrich only
• ontology_mode 'strict': drop unmatched entities; an empty or missing ontology file is a hard error
• Default loader order: code, text, pypdf, image, audio, video, dlt_csv, csv, unstructured, advanced_pdf, docling

Safety notes: network, scripts and sensitive data

This skill does not run scripts. It does use the network. URL fetching is allowed by default through ALLOW_HTTP_REQUESTS=true. Git cloning of repositories and database connections read or send data outside the local machine. Treat each of these as a step that needs deliberate approval, especially for sensitive data. Review what the agent plans to fetch or connect to before it runs.

The open-source GLiNER extractor is a demo with lower accuracy, according to the skill's own file. Do not rely on it for production graph extraction.

• Uses network: yes (URL fetching, git clone, database connections)
• Runs scripts: no
• Review sources before the agent ingests URLs, repositories or databases

Requirements and optional extras

You need Python with the cognee library installed. LLM-based extraction needs LLM_API_KEY. GLiNER does not need it. Repository ingestion needs git on PATH. Session writes with session_id need CACHING=true. The GLiNER option is an optional install, and its model is about 750 MB and downloads on first use.

• Python with the cognee library installed
• LLM_API_KEY for LLM extraction (not needed with GLiNER)
• git on PATH for repository ingestion
• cognee[gliner] extra: about 750 MB model download on first use
• docs or docling extra for Office formats
• CACHING=true for session writes

Fix changed files and keyword argument errors

Re-remembering the same path with different content raises DocumentUpdateRequiredError, because it is an update, not a new document. Use cognee.update(data_id=..., data=..., dataset_id=...), which re-extracts only the changed chunks and keeps the document's id. Identical content is a no-op.

remember() forwards keyword arguments through a fixed allow-list. Anything else raises TypeError: Unexpected keyword arguments. Some real options from lower-level functions are not on remember() yet, so check the workaround before you assume a feature is missing.

• Changed file: use cognee.update() instead of re-remembering
• Unexpected keyword argument: check the allow-list and the workaround for the option
• extraction_rules works through preferred_loaders with beautiful_soup_loader but needs the scraping extra