| name | knowledge-graph |
| description | Index curated reference corpora into a searchable knowledge graph via the cognee engine, then query it to ground technical writing, review, and analysis. Use when ingesting reference documents into named datasets or retrieving grounding context for tasks like drafting or reviewing a manuscript. A reference-knowledge index, explicitly not agent session memory. |
Knowledge graph for reference grounding
Mental model
Treat cognee as a searchable, semantically-indexed reference library built over a corpus you curate by hand.
You ingest chosen reference documents into named datasets, build a knowledge graph across them, and query that graph for grounding context when writing or reviewing sophisticated technical material.
The 1.0 API verbs are named for memory (remember, recall, improve, forget), but here they index and query a reference corpus, not session state.
Despite the names, remember ingests reference docs and builds the graph, recall queries it for grounding, improve enriches it, and forget deletes data.
The discipline lives in what you ingest (curated references) and why you query (grounding), not in avoiding the verb names.
The dataset is the boundary of connected reasoning: documents within one dataset link into a single connected graph, while separate datasets are (under the default access-control mode) isolated graphs that multi-dataset search only unions, so co-locate material you want reasoned over together — see references/datasets.md.
This engine has zero relation to AI-agent session memory: it does not store conversations, reasoning traces, or chat history.
Throughout this skill and its references, retrieved context is evidence indexed from source documents, and load-bearing claims must be verified against the originals.
When to use, when not
Reach for this skill when you are populating a reference dataset or pulling grounding context for a task such as drafting or reviewing manuscript.qmd.
Do not use it as session memory.
Three anti-patterns to reject: asking cognee to "store this conversation so it remembers next session"; auto-capturing tool calls or reasoning traces into a dataset; and using recall -s <session-id> to fetch prior chat history.
Those are the memory use we explicitly do not adopt; the session-Q&A surface (sessions, feedback) exists but is not the reference-grounding focus.
Core lifecycle
The everyday path uses remember (ingest plus build plus self-improvement) then recall.
For bulk corpus loads, decouple ingest from build: many add calls, then one cognify, then search.
The wrapper is invoked bare as cognee and targets the SaaS platform; verify a verb's flags with cognee <verb> --help (deployed pin is tag cognee-v112).
cognee remember ./refs/*.md -d engineering-references
cognee recall "what does the corpus say about X" -d engineering-references
cognee add ./refs/part-a.md -d modeling-references
cognee add ./refs/part-b.md -d modeling-references
cognee cognify -d modeling-references
cognee search "grounding question" -d modeling-references
References
| Reference | Covers |
|---|
| references/cognee-cli.md | running cognee — all 14 verbs, flags, raw cognee-cli escape hatch |
| references/datasets.md | dataset naming and boundaries, lifecycle, hygiene |
| references/search-types.md | choosing a recall/search retrieval mode (-t) |
| references/ingestion-workflow.md | populating engineering-references / modeling-references end-to-end |
| references/architecture.md | ECL pipeline, engine model, remember = add + cognify + self-improvement, when to decouple |