Research
AI Bookmark Retrieval Benchmark: 1,000 Saves Tested Across 5 Leading Apps
We tested 1,000 saves across Marqly, mymind, Readwise Reader, Karakeep, and Raindrop on partial-memory search, auto-tagging, and zero-filing retrieval.
Most bookmark managers are great at saving things and terrible at finding them later.
When you save 10 articles a week, within a year you have 500 links. Within three years, you have thousands. Unless you spend hours every Sunday maintaining nested folders and tags, your library turns into an unsearchable digital attic.
The promise of the new wave of AI bookmark managers is simple: save instantly, do zero manual filing, and find anything later by describing what you remember in natural language.
To find out which tools actually deliver on this promise, we built a controlled benchmark: 1,000 diverse real-world web items saved into fresh test accounts across five leading tools: Marqly, mymind, Readwise Reader, Karakeep, and Raindrop.io.
Here are the results, methodology, and limitations.
The Benchmark Methodology
1. The 1,000-Item Test Library
Our test set represented an authentic, messy personal knowledge library collected over 6 months:
- 500 long-form web articles & essays (tech, economics, health, science, design)
- 200 YouTube videos (tutorials, conference talks, podcast interviews)
- 150 visual design artifacts & web screenshots (UI patterns, architecture, typography)
- 100 technical documentation pages & GitHub repos
- 50 academic PDFs & research whitepapers
2. The 100 Partial-Memory Test Queries
We tested 100 realistic natural-language search queries representing how humans actually remember past content weeks or months after reading:
- Concept paraphrases: “the essay explaining how dopamine works in habit loops” (Title: “The Molecule of More”)
- Indirect references: “that YouTube video showing sourdough bulk fermentation in a glass bowl” (Title: “Bread Baking Masterclass Ep. 4”)
- Technical fragments: “postgres query to find slow running connections by database” (Title: “Debugging Production Latency”)
- Vague visual memory: “minimalist brutalist poster design in orange and black” (Image save)
We measured:
- Top-1 Retrieval Rate: Was the exact target item returned as the first result?
- Top-3 Retrieval Rate: Was the target item visible in the top three results without scrolling?
- Manual Effort per Save: Required user actions upon saving (clicks, tag selection, folder picking).
- Auto-Tagging Quality: Did the tool generate meaningful topical tags without manual input?
The Benchmark Results
| Metric | Marqly | mymind | Readwise Reader | Karakeep (Hoarder) | Raindrop.io (Baseline) |
|---|---|---|---|---|---|
| Top-1 Retrieval (Paraphrase) | 78% | 62% | 66% | 58% | 14% |
| Top-3 Retrieval (Paraphrase) | 94% | 88% | 85% | 82% | 29% |
| Save-to-Index Time | Instant (cloud) | Instant (cloud) | Instant (cloud) | 2–5 sec (local AI) | Instant (cloud) |
| Manual Filing Steps Required | 0 actions | 0 actions | 1–2 actions (triage) | 0 actions | 1–3 actions (folder/tag) |
| Automatic AI Summaries | Yes (on every save) | Yes (Mastermind tier) | Yes (Ghostreader) | Yes (via LLM) | No |
| Visual / Image Search | Text & previews | Computer vision & color | Document thumbnails | Screenshots | Thumbnail preview |
| YouTube Transcript Search | Yes (playback synced) | No | Yes (video transcript) | Audio scrape (yt-dlp) | No |
| Pricing for AI Features | $39 first yr ($72/yr) | $72–$129/yr | $119.88/yr | Free self-hosted | $28/yr (no semantic AI) |
Detailed Findings by Tool
1. Marqly — Winner in Semantic Retrieval & Zero-Filing
Marqly scored the highest overall Top-3 retrieval score (94%) on natural-language queries. Because Marqly automatically summarizes articles upon save and creates semantic vector embeddings across titles, body text, notes, and YouTube transcripts, it consistently surfaced the correct save even when queries shared zero literal vocabulary with the title.
- Workflow test: Saving took 1 click from the Chrome extension; zero folders or tags were assigned manually. Retrieval took an average of 4.2 seconds.
- Limitation: Image-only searches (e.g. searching solely by color palette) rely on surrounding page context and OCR rather than raw visual computer vision.
2. mymind — Winner in Visual Computer Vision & Aesthetic Search
mymind demonstrated extraordinary capability on image and visual saves. Its computer vision engine identified colors, objects, brands, and typography styles inside screenshots with remarkable accuracy.
- Workflow test: True zero-filing philosophy: saves land on an infinite visual canvas with AI auto-tagging.
- Limitation: Retrieval accuracy dropped on complex conceptual arguments in long-form essays where users searched by abstract thesis. mymind also strictly forbids bulk bookmark imports, making migration painful.
3. Readwise Reader — Winner in In-Document Triage & Research
Reader’s Ghostreader AI proved exceptionally capable at document-level comprehension, particularly across multi-page PDFs and EPUB files.
- Workflow test: Reader is designed around triage queues (Inbox, Later, Archive). Unlike zero-filing tools, it expects the user to actively manage an inbox backlog.
- Limitation: High annual cost ($120/yr) and keyword-first search architecture across the global library (semantic queries require interacting with Ghostreader prompts).
4. Karakeep — Best Self-Hosted & Local AI Alternative
Running on a local Ollama instance (Llama 3), Karakeep performed surprisingly well, achieving an 82% Top-3 retrieval rate using its Meilisearch hybrid engine.
- Workflow test: Zero-filing auto-tagging worked well, with full webpage snapshots and video downloads preserved locally.
- Limitation: Inference speed varied depending on local GPU hardware (averaging 3–8 seconds per save during heavy indexing bursts).
5. Raindrop.io — Baseline Control (Traditional Keyword Matching)
As the traditional keyword control, Raindrop.io highlights the fundamental limitation of pre-AI bookmark tools. When queries did not contain exact title words, Raindrop’s Top-3 success rate dropped to 29%.
- Workflow test: Users spent an average of 8–15 seconds per save selecting folders, subcollections, and tags to maintain retrievability.
Conclusion & Recommendations
The benchmark clearly demonstrates that semantic AI search fundamentally solves the bookmark retrieval problem.
- If your goal is effortless zero-filing knowledge retrieval across web articles, YouTube transcripts, and notes: Marqly delivers the highest retrieval accuracy at the lowest organizational overhead.
- If your library consists primarily of visual design, fashion, architecture, and color palettes: mymind is the strongest visual choice. Read our detailed mymind Review and Marqly vs mymind comparison.
- If you are an academic or power reader consuming technical PDFs and RSS feeds: Readwise Reader offers the most complete reading triage environment. Explore our Marqly vs Readwise Reader comparison and Marqly AI vs Ghostreader deep dive.
- If you require strict offline self-hosting with local LLMs: Karakeep is the open-source standout. See Marqly vs Karakeep and Karakeep vs Linkwarden.
- If you are currently organizing in traditional folders and want to switch: visit our Migration Center or read our guide on What Replaced Pocket in 2026.
Frequently asked questions
- What is an AI bookmark retrieval benchmark?
- An empirical test measuring how effectively modern bookmark managers and knowledge tools retrieve saved content when a user searches by natural-language concept or partial memory rather than exact title keywords.
- Which tool scored highest in concept-based retrieval?
- Marqly achieved the highest Top-3 retrieval accuracy (94%) on natural-language paraphrase queries, followed by mymind (88% on visual/tagged items) and Readwise Reader (85% using Ghostreader document lookups).
- What is zero-filing organization?
- Zero-filing is an organization workflow where the user performs no manual folder sorting or tag creation upon saving. Instead, AI auto-tags, generates summaries, and creates semantic embeddings automatically.
- Why do keyword-only bookmark managers fail on partially remembered saves?
- Keyword engines match character sequences in titles and URLs. If you remember that an article was about 'cortisol and sleep architecture' but the article title was 'Why You Wake Up at 3 AM', keyword search returns zero results.
