Tools for Thought
Tools for thought are technologies, notations, and practices that expand the range of thoughts a person is able to think. The phrase names both a philosophical tradition about human cognition — older than computing itself — and a modern category of software descended from mid-twentieth-century research into intelligence augmentation. Howard Rheingold's 1985 book Tools for Thought: The History and Future of Mind-Expanding Technology gave the field its name and traced its lineage from nineteenth-century logicians through the pioneers of interactive computing.
The intellectual history runs from Vannevar Bush's 1945 essay "As We May Think" and its memex, through Douglas Engelbart's vision of augmenting human intellect and the 1968 demonstration of his oN-Line System, to Ted Nelson's coining of hypertext, Alan Kay's Dynabook, and the graphical interfaces built at Xerox PARC. Two rival conceptions run through this history: intelligence augmentation, which designs computers to extend human judgement, and artificial intelligence, which designs them to act autonomously. The same tension reappears in human-computer interaction as the choice between direct manipulation and interface agents. A 2026 conference paper argues that the term has itself drifted, undergoing a quiet but significant change in meaning since its origins in the 1980s[^c12]: from Iverson's 1976 lecture on notation as a tool of thought, through Rheingold's account of computers generally, to its re-cooption around 2018 to describe Markdown-centric, wiki-linked, web-based noting tools — a revival the paper argues unknowingly rediscovered ideas long familiar in hypertext research.
The tradition's core ideas became practical technologies: the hyperlink and the World Wide Web, the wiki, and eventually bidirectional links, transclusion, and graph views. Contemporary tools for thought — Obsidian, Roam Research, Logseq, Notion, Heptabase, TiddlyWiki, and Org-mode — implement methods such as the Zettelkasten, Building a Second Brain, spaced repetition, linking and backlinks, outlining, mind mapping, and progressive summarization. Most store notes as plain text or Markdown files, a choice justified by durability, portability, and, increasingly, compatibility with language models. Hypertext research has also renewed one of its founding problems — deriving navigable structure from a corpus that arrives without it — with a 2026 paper that induced guided trails through a flat collection of more than 300,000 questions and then scheduled revisitation with an FSRS-7 spaced-repetition scheduler at concept granularity[^c36].
The field is also a subject of debate and, since the mid-2020s, of organised research. Advocates argue that externalised, linked knowledge compounds and improves thinking; critics note that the evidence for specific personal-knowledge-management practices is thin, that many systems are abandoned, and that tooling can become a form of productivity theater. Two developments have sharpened the debate. First, generative AI has been positioned as both a risk to cognition and a means of protecting it. A Tools for Thought workshop at the CHI conference mapped the resulting questions in 2025 — convening 56 researchers around 34 papers on metacognition, critical thinking, memory, and creativity — and its synthesis, a growing special issue of the journal Human-Computer Interaction, and a follow-up workshop in 2026 now define a research agenda around outcomes, strategies, and adoption. Design research within this programme has begun to separate forms of augmentation experimentally: a CHI 2025 study found that an assistant which extended users' own decision-making rationales integrated better into their thinking and produced slightly better outcomes than a recommendation-based assistant[^c4], while a first-party Microsoft account framed the goal as designing shared cognition rather than answering questions, warning that over-automation yields documents without authorship and a crisis of thought rather than a productivity crisis[^c3]. A 2026 study pushed the argument further by deliberately introducing friction, suppressing a dominant concept inside a language model so that it could not simply supply the obvious answer; it found that the unmodified model reduced mental demand and effort but also reduced perceived ownership, while the constrained model produced ratings closer to unaided ideation[^c13] — evidence for a stage-sensitive view in which assistance should preserve effort during early ideation. A 2026 position paper extended the design argument from cognitive friction to affect, arguing from enactive cognitive science that affect is constitutive of cognition, reshaping the trajectory of thinking rather than merely its speed[^c28], and proposing shared attention and affective reorienting as design goals for AI-supported thinking tools. Second, AI has begun to change what a knowledge base is for: rather than retrieval scaffolding for human memory, some practitioners now design "typed memory" that an agent can read, update, and act on directly, with assistants such as personal AI agents described as "second brains." That aspiration also undercuts the manual linking and graph-viewing those tools once encouraged, shifting the site of design from the link to the check, the rule, and the curation step that keeps a human in the loop. The industry framing of the gap is blunt: an influential 2026 essay argued that knowledge software is built for making rather than thinking[^c5], leaving no modern exploration tools outside the language models themselves, which have emerged as thinking partners[^c6]. A parallel design response makes the assistant deliberately non-directive — NEC demonstrated an educational "Thinking Support AI" that answers students' claims with questions rather than answers[^c8], and a 2026 consumer product positioned itself as a private thinking partner that reflects a person's reasoning back and leaves the conclusion to them[^c7].
By late 2026 the commercial landscape had moved in several directions at once. Leading workspaces such as Notion and Tana exposed interchangeable frontier models and connected to external agents through the [[model-context-protocol|Model Context Protocol]]; local-first applications such as Reflect rebuilt themselves around open-source, plain-Markdown files to stay legible both to people and to AI; a distinct category of AI memory layers — Zep, Mem0, Letta, and others — emerged to give agents persistent state across sessions; and organisations began routing expert-validated knowledge into the same assistants their staff already use. Portability became a product category of its own: Egoist Machines launched AI Passport, a permission-based layer that carries a user's context across ChatGPT, Claude, Gemini, and Perplexity on the principle that a user's context should sit with the user and applications should come asking for it[^c25]. Notion extended itself from a workspace into a programmable agent platform, with customers reported to have built over one million agents on it by mid-2026[^c14], and then deliberately positioned itself less as an agent than as the context layer agents run on, arguing that becoming the best context layer is what unlocks automation and opens the path to autonomous AI teammates[^c29] — a strategy its growth bore out, with enterprise annual recurring revenue up 169 percent year over year as of August 2026 and two-thirds of that revenue attributable to AI involvement[^c30]. Google took the opposite route, releasing Gemini Agent as its first workplace personal agent, provisioned with an enterprise's business context and able to answer questions, create content, and write code from a single prompt box[^c31], while a further entrant targeted teams rather than individuals: Scuba raised $34 million for a shared context engine, on the founders' argument that frontier models are not intelligence bound but context constrained[^c32]. China's Kingsoft launched WPS Notes, an AI-native, full-modality cloud notebook built around automatic multi-level tagging and a Model Context Protocol server[^c15]. Productivity suites also consolidated around AI note-taking and agents, as Superhuman acquired the meeting assistant Fathom and Meta connected its Muse agent to third-party knowledge tools including Notion and Granola, while dedicated capture devices, grouped under the emerging category of [[ai-meeting-note-takers|AI meeting note-takers]], promised note-taking without retained recordings. Note-taking itself moved toward on-device processing and agentic production: Google shipped a Mac notetaking app built on a local embedding model that organizes, searches, and connects information directly on consumer hardware[^c9], while Plaud organised its platform around an agent that, given a task, sees it through[^c10] and Tanka AI built a team "memory" by combining brain-science-derived long-term memory technology with large language models[^c11]. Two further systems automated the construction of the knowledge base itself: Sentient OS, an on-device agent that reads a user's email, files, and chat records during idle overnight time to generate a personal knowledge base without manual filing[^c34], and openJiuwen's personal knowledge space, which collects from configured sources and organises them into linked, sourced Markdown knowledge that an agent draws on with progressive disclosure[^c33]; against that grain, the open-source agent nanoMuse stored its memory as files the person can read and left the model as the user's choice[^c35]. The networked note-taking camp continued to shift as well: Logseq invited its monthly backers to test the database beta and real-time collaboration sync it funds through Open Collective[^c16], even as its file-based edition settled into maintenance mode as "Logseq OG"[^c37], and a broad "vibe notetaking" consensus formed around a single pattern — the user dumps unstructured content and the AI handles indexing, tagging, and linking, so that organisation becomes an emergent property of AI processing rather than a human discipline[^c17]. Even long-running training in the field has been unsettled: Harold Jarche announced that his Personal Knowledge Mastery workshop series would end after its final cohort beginning 5 October 2026[^c2], citing the rise of generative AI among his reasons[^c1]. A parallel category of AI notebooks instead answered questions over a corpus the user supplied rather than over the user's own files: Google's Gemini Notebook (renamed from NotebookLM in July 2026), whose 2026 upgrade gave each notebook a secure cloud computer able to execute code across more than 100 software skills, and Microsoft's Copilot Notebooks, whose 2026 upgrade reframed the notebook from a retrieval library into a workspace that delegates tasks on a schedule, so that digital notes become intelligent partners capable of proactively helping carry out tasks[^c27]. Underneath the churn, the field's central question became architectural as much as methodological: whether an assistant can reach a person's notes at all, and whether it may write back — the question is not which program draws the prettiest links, but whether an AI can reach those notes at all, and whether it is allowed to leave anything behind[^c26].
The evidence about what generative AI does to thinking has itself become a live question. Reviews published in 2026 converge on the view that the outcome depends on the mode of engagement rather than on use as such: a systematic review distinguishes passive AI consumption, assisted AI learning, and reflective AI collaboration as producing markedly different cognitive results[^c19], while a critical review argues that AI chatbots redistribute rather than diminish cognition, shifting effort away from offloadable tasks and toward evaluation, synthesis, strategic direction, and metacognitive oversight[^c18]. The field has also begun to enter university teaching, with courses that treat cognitive augmentation as a continuous tradition whose latest stage is collaboration with AI[^c20], and design research has proposed treating a system's refusal to generate text as a core feature rather than a missing one[^c21].
Outside knowledge work proper, the question of whether AI extends or supplants human reasoning became a live test case in research mathematics. In October 2026 Meta AI published six mathematics papers co-authored with its Muse Spark system, five of them touching problems that were previously open, and Google DeepMind's AlphaProof Nexus — which verifies its reasoning with the Lean proof assistant — reportedly solved nine previously unsolved Erdős problems, results published in Science alongside commentary declaring that the era of AI for research-grade mathematics had arrived[^c22]. Commentators framed the ideal division of labour as AI expanding the space of conjectures to explore while humans decide which explorations are worth keeping[^c23], even as some mathematicians warned that large-scale use of AI could weaken the human drive to understand proofs. A parallel programme tested whether language models can remove the burden of writing the prompts that spaced-repetition memory systems depend on; a 2026 report found that even the strongest model it tested produced unusable prompts roughly a third of the time, and that instruction, fine-tuning, and reinforcement learning improved efficiency rather than capability[^c24]. Both cases restate the field's founding tension: the same capability that offloads a difficult human task also risks displacing the effort through which the task is understood.