Adding Content

Your Cortex is only useful if you put things in it. Graphnosis supports several ingest methods: files, URLs, and quick clips.

Supported file types

Format	Extension(s)	Notes
PDF	.pdf	Parsed via pdfjs-dist in a worker thread; text + structure extracted
Word document	.docx	Text and basic structure extracted
Markdown	.md, .mdx	Full text; frontmatter treated as metadata
Plain text	.txt	Ingested as-is
HTML	.html, .htm	Main content extracted; nav/footer/scripts stripped
JSON	.json	Stringified and chunked
CSV	.csv	Rows converted to text blocks

Binary files (images, audio, video, executables) are not supported and will be rejected at ingest.

Add a file

From the menu bar: Click the Graphnosis icon → Add File → pick a file.

By drag and drop: Drag a file onto the Graphnosis menu bar icon.

From the main window: Open the Graphnosis window, go to your graph, and click + Add.

After you pick a file, Graphnosis will:

1. Parse and extract text content.
2. Split the text into overlapping chunks (default: 512 tokens per chunk, 64-token overlap).
3. Embed each chunk locally using BGE-small-en-v1.5 (ONNX, runs in a forked child process).
4. Encrypt the chunks and embeddings and write them to cortex.db.
5. Create a Source record linking the chunks back to the original file path.

Ingestion of a large PDF (hundreds of pages) may take 10–30 seconds on Apple Silicon, longer on older Intel hardware.

What is a Source?

A Source is a record that represents one piece of ingested content — a file, a URL, or a clip. It stores:

• The original location (path or URL)
• The ingest timestamp
• The graph it belongs to
• Chunk count and total token estimate
• A SHA-256 content hash (used for deduplication)

If you ingest the same file twice, Graphnosis detects the duplicate hash and skips it unless you force-reingest.

Ingest a web page

Click the Graphnosis icon → Add URL → paste a URL.

Graphnosis fetches the page, strips boilerplate (navigation, ads, scripts), and ingests the main article body. JavaScript-heavy single-page apps may not extract well if the content isn’t in the initial HTML.

Save a clip

A clip is a short piece of text you paste or type directly — a quote, a decision, a note, anything that doesn’t have a file.

Click the Graphnosis icon → Add Clip → type or paste your text → choose a graph → Save.

Clips are useful for:

• Capturing decisions made in a meeting
• Saving a snippet from a conversation that isn’t worth a full file
• Quick notes you want to be able to recall semantically

Large files and chunking

Files are split into chunks before embedding. The default chunk size is 512 tokens with a 64-token overlap between adjacent chunks. This overlap helps recall find content that falls near a chunk boundary.

For very large files (thousands of pages), ingest is done in a background worker so the UI stays responsive. You can see ingest progress in the Graphnosis window.

Who added each memory? AI client attribution

Memories added by an AI client (via the remember MCP tool — Claude Desktop, Cursor, Claude Code, etc.) show a small turquoise badge in the Sources list:

Badge	What it means
via claude-ai	Claude Desktop ran remember
via cursor	Cursor ran remember
via claude-code	Claude Code ran remember
(no badge)	You added this manually — drag-drop, paste, or file picker

The badge derives from the MCP initialize handshake’s clientInfo.name. It’s purely informational — there’s no behavior difference; the badge just helps you audit what came from where.

Two flavors of AI-saved memory: the remember tool now accepts kind: 'clip' | 'ai-conversation'. AI clients use clip (the default) when extracting a fact from external content (a doc you shared, a search result, an article). They use ai-conversation when saving a turn or summary of the CURRENT conversation — so you can tell “Claude paraphrased me” from “Claude saw this in a doc I shared”. Both appear in the Sources list; the source ref makes the kind explicit.

Correction attribution: when an AI client applies a correction via the apply MCP tool (after you approved a correct proposal), the op-log records correctedBy: <client name> on every node edit/supersede/delete. The Activity view surfaces this for full audit.

Auto-relink: how new memories connect to old ones

After every ingest, Graphnosis runs a quick auto-relink pass over the affected engram. It compares the entities mentioned in the newly-added nodes (people, places, projects, concepts) against the entities in every existing node and creates “related-to” edges wherever there’s meaningful overlap.

This is why you’ll see lines like this in the dev terminal during normal use:

[host] auto-relink wove 44 edges across 373 active nodes in personal

That’s Graphnosis discovering, on its own, that a new note you just dropped in has 44 cross-references to memories you already had. Without this pass, each new file would be an island; with it, your engram graph genuinely gets denser the more you put in.

The pass is throttled and capped (autoRelinkMaxNodes in Settings, default 5,000 active nodes) so it stays fast even as graphs grow. For very large engrams that exceed the cap, you’ll see auto-relink skipped: active node count > maxNodes — recall and search still work; only the cross-doc edge creation is paused.

Content caching: why it matters for recovery

When you ingest a file, the raw bytes are also cached in your Cortex (encrypted, in content/) — by default for any source up to 512MB. This is the difference between:

• Cache hit: if you later move/delete the original file, or the .gai file is somehow damaged, you can rebuild from cache without ever touching the original source again. Fast, automatic, no user action beyond clicking “Recover” in the app.
• Cache miss (e.g. you raised the cap, or disabled caching): recovery only works if the original file is still at the exact path you ingested it from. If it’s moved, the source shows up as file-missing in the Recovery panel.

You can change the cache mode and per-source size limit in Settings → Content Cache if you want to trade disk space for recovery flexibility. The default (all mode, 512MB cap) covers most practical sources including large reference manuals.

There is no hard file size limit, but files over ~50 MB of raw text may take several minutes to ingest.

Reingest and update

If a file you’ve ingested has changed, you can reingest it from the Source detail view. The old chunks are replaced with fresh ones from the updated content. The Source’s updated_at timestamp is updated accordingly.

Removing a source

Open the Source in the Graphnosis window and click Remove. This deletes all associated chunks and embeddings from the Cortex. The original file on disk is not touched.