@Vlad#1, you've invested in The Assistant, but you have not implemented subsystems for the cybernetic closure of Kagi with LLMs. The Assistant still processes information at speeds higher than human capabilities, despite that as of February 2025 we remain biologically limited to a historical baseline in available compute.
Recalling and then retyping previous Assistant outputs into the system repeatedly created a cognitive bottleneck, because the user must reconstruct information, costing both time and opportunities to use valuable data when it is most needed. Although keywords could be considered a partial workaround, they offer don't provide an throughput: even if you manage to recall the keywords, you still need to search for the exact piece of information you from that conversation & copy-paste it manually. Then, imagine your life in an alternative world where human thought flows seamlessly with the computational system. To achieve that, you need to bridge the gap of implementation of The Library.
The foundations of The Library will based on UTF‑8 text files called stacks, which function as a storage & a temporal interface for the user's interactions within Kagi. Each stack is editable, downloadable, and uploadable, partially mimicking note‑taking software without enforcing a particular syntax on the user.
Its primary strength, however, lies in the integration with the pre‑existing Kagi system—how it interfaces Quick Answers, Search Summaries, various bookmarking tools, the “Ask questions about this site” feature, outputs of The Assistant's through The Lens.
By pushing information from Quick Summaries, Site Summaries, Document Discussions, or Assistant messages to designated stacks, the user can create a persistent "worksphere" that extends their memory. This worksphere is accessible through Lenses, thereby impacting both Kagi Search and The Assistant. When creating a Lens, the user selects—along with current options—relevant stacks from the Library, preloading context for future queries.
For instance, consider these examples:
Site Summary ⇒ The Lens ⇒ The Assistant
A user may save summaries of dad jokes from various sites to a Dad Jokes stack. Later, on a friend's birthday, they could ask The Assistant to recall the most relevant birthday‑related jokes from that stack and adjust them for the greeting card.
Similarly:
• A person with style might maintain an Inventory/Accessories stack, an Inventory/Garments stack, and an Inventory/Makeup stack, based on searches for items they have historically purchased. A Wardrobe Custom Lens would then aggregate these and feed them to a custom Assistant named Private Stylist.
Document ⇒ The Assistant ⇒ The Lens ⇒ Quick Answer
In another instance, the user may upload a PDF to The Assistant and, at the end of a discussion, save an LLM wrap-up containing promising business contacts into a stack called Business Networking. A quarter later, when they need these contacts, they type in the browser search bar:
"past redistributors who bought toothbrushes from us & have higher than 1M revenue?"
Using a bang command (e.g., !bcontacts), they obtain the Kagi Quick Answer required for professional use. After reaching out, this information can be passed back to The Assistant and saved in a Business Contacts stack, closing the loop—the only copy-paste interaction being the transfer of contact information from Quick Answer to an email client.
Similarly:
• A sales manager might save summaries of client meeting notes in a Client Insights stack and later retrieve the necessary details when an immediate need arises during a follow‐up call.
• An entrepreneur could store key financial reports in a Financials stack and later request a summary when an official inquires about the details.
The Assistant ⇒ The Assistant
Consider everyday personal use: compiling LLM‑generated recipes into a Recipebook stack, travel itineraries into a Travel Plans stack, and instructions for proper exercise form into an Exercises stack—or perhaps saving playlists of music exported from external streaming services into text files. When planning a party, vacation, or workout schedule, the user could ask The Assistant to seamlessly recall the appropriate content.
The cybernetic closure (in the broader sense of the word) unlocks an entirely new level of efficiency, freeing human cognitive resources for far more pleasant tasks than painstakingly recalling exact details from an LLM months ago.
────────────────────────────
Side details about integration with The Lens:
• Versioning text stacks semantically—for example, naming a file notes@1.2.3.txt—would provide a smoother experience.
• Saving to a stack should be executed in a UX mode akin to “instant fire‑and‑forget.”
• Lenses can dynamically inject relevant stacks into the query context because Kagi’s hybrid search is neural network‑based; even with fuzzy or incomplete human recall, relevant results can still be fetched.
• Integration with Lenses ensures scoping capabilities so that a user who does not wish to mix online search results with those from stacks can avoid doing so. This would be particularly useful, for example, when a user is writing a Assassin's Creed or X‑Files fanfiction, and does not want Kagi results to constantly suggest that they believe in conspiracy theories.
────────────────────────────
On the worst‑case time cost of human recall compared to now‑obsolete alternatives:
The purpose of knowledge management software is to help with recalling information. It is reasonable to assume that users may forget something. When they forget keywords, retrieving information from a standard notebook (e.g., Notes, Joplin) typically takes linear time, as you must search through portions of it. A wiki‑style system (e.g., Obsidian*) may take logarithmic time, since recalling an adjacent concept might suffice and linking or tagging is required. In contrast, my system enables human information retrieval in constant time.
────────────────────────────
Extension (extended):
Over a 2–3‑year time scale, consider a topology where everything about a user is stored within a virtualized Linux system, with temporary passwordless accounts providing The Assistant with real‑world agency, since open‑source autonomous browsing will likely become competitive with proprietary technology. It would be trivial to add apps like browsers, commands, or permissions to The Lens as options (with LLMs informed via wrappers), though you would probably want to restrict LLMs from recursively calling other Kagi‑provided LLMs ad infinitum. Structurally, this resembles Langchain, although in my view each user account would map to a node in a graph, with commands/permissions mapping to edges. The LLMs for such accounts would not need to be local, since Kagi can run a system package that delegates the load to inference providers. For example:
kagi spawn agentic‑assistant --model=o3 username=indie‑game‑development [Make Tetris in Python, including verification logic, and use the command line to execute it].
────────────────────────────
Economically:
I admit that aside from the direct editing of stacks—a stack‑like interaction—the UX may not be optimal. Nonetheless, after implementing The Library, you would:
- Offer a competent alternative to Claude Pro for professional customers.
- Automate context management by eliminating the need to link individual elements—thereby addressing PKM usage friction.
- Reduce churn by cognitively augmenting existing users.
I could not devise a more lucrative extension of the Kagi system, which means above would probably work.
*Obsidian LLM plugins exist, but at the time of writing, Kagi's was technologically ahead; for example, Hades can summarize inputs of unlimited size.
