SMF·RECALL
Can a prompt replace Recall?
Read it later & RSS — read-later, bookmarks and RSS
Exhibit tracking slip
Verdict
The saving and summarising half is ordinary work. The graph is what makes Recall distinctive, and building it well means resisting the obvious approach: linking by embedding similarity alone produces a hairball where everything connects to everything. Extracting named entities and linking on shared entities gives a graph with meaning in it, which is a sitting's more work and the difference between a useful tool and a pretty picture.
Exhibit A — The prompt
Received on31.07.2026Build a personal knowledge graph fed by things you save, where links come from shared entities rather than raw similarity.
Sources: a web page by URL, a YouTube video by URL (using its transcript), a PDF uploaded, and pasted text. Fetch politely, honour robots.txt, run Readability for web pages, and never attempt to get past a paywall or a login.
Extraction, per saved item:
- A summary of a few sentences, and three to five key points.
- Named entities, typed: people, organisations, products, places, and concepts. Ask the model for these as a structured list, and normalise each one — case-folded, common suffixes stripped, and matched against entities already in the graph so "OpenAI" and "Open AI" become one node rather than two.
- The item's own topic, chosen from the user's existing topic list where one fits, and proposed as new only when nothing does. Uncontrolled topic creation is what turns a graph into noise.
The graph: items and entities are both nodes. An item links to every entity it mentions. Two items are related when they share entities, weighted by how rare the shared entity is across the whole collection — a shared mention of an obscure researcher means far more than a shared mention of a large company that appears in half your library. Compute this weight explicitly and show it, so a surprising connection can be explained.
Deliberately do not link on embedding similarity alone. It produces a graph where everything is loosely connected to everything, which looks impressive and answers no question. Embeddings may be used as a secondary signal for the related-items list, clearly labelled as such.
Views: an item page with its summary, key points, entities and related items ranked by shared-entity weight. An entity page listing every item mentioning it, chronologically, which is the view that turns the graph into research. A graph visualisation, deliberately filtered to one entity's neighbourhood rather than showing everything at once.
Ask: answer a question from the collection by retrieving the highest-weighted items for the entities in the question, sending only those, and requiring the answer to cite the items it used, with a plain refusal when the collection does not contain an answer.
Storage: PostgreSQL for items, entities and edges. Full-text search across summaries and original text.
Use Anthropic for extraction and answering. With no key, saving, storing and searching must still work, with entity extraction disabled and clearly marked.
Out of scope: a browser extension, mobile apps, sync, and any recommendation of content you have not saved.Opening prefills the prompt — press enter to run it.
Exhibit B — What you lose
- B.1 the browser extension and mobile capture
- B.2 sync between devices
- B.3 hosted extraction that handles awkward sources reliably
- B.4 the polish of a graph view built for exploration
Prior art
Exhibit C — Why people still pay: capture polish, sync, and content partnerships
Because capture has to be frictionless or the graph stays empty, and a knowledge base with thirty items in it is not a knowledge base.
Questions
Can I import my Recall library?
Recall exports saved items with their summaries, and the URLs and text come across. The graph does not export, but it does not need to — the whole point is that entity extraction rebuilds it from the items on import, which takes one model call per item.
Why not just link by embedding similarity?
Because it produces a graph where every item is mildly related to every other, which is visually impressive and useless. Shared-entity linking weighted by rarity gives edges you can explain, and being able to answer "why are these two connected" is what makes the graph worth having.
What does it cost to run?
One extraction call per saved item — summary, key points and entities in a single structured request — which is a fraction of a cent for an article and a few cents for a long PDF. Saving a hundred items a month costs about a dollar, plus a small VPS.
What is the one thing that does not survive the rebuild?
Frictionless capture. A knowledge graph is worth exactly as much as what is in it, and without a browser extension and a phone share sheet you save the things you remember to paste in — which is a fraction of what you actually read.
Related tools
Receipt