Second Brain vs. AI Assistant: How to Integrate Both Without Losing Your Own Thinking

TL;DR. A notes system does two jobs: it stores information, and it builds memory and judgment in your own head. An AI assistant is excellent at the first and, used carelessly, removes the second. Generating material yourself beat reading it across 86 studies (effect size .40) and by about 10 percentage points across 310 experiments. Testing yourself beat restudying across 159 effect sizes (g = 0.50, and 0.73 with feedback). Saving a file to a trusted place improved memory for the next file, while offloading during a task boosted performance but diminished memory in three experiments of 172 people each. The “Google effect” priming result did not replicate: the pooled replication effect was 0.135 of the original. CEO move: design the system, with written roles, AI read and write rules and a weekly 20-minute review, instead of buying another app. Student move: keep the generative steps (writing in your own words, deciding links, answering review questions from memory) and give AI storage, search, sorting, and feedback after your attempt.

This piece assumes you already give your assistant standing context; if not, start with the context file and the guide to custom instructions, projects and memory. For the wider offloading debate, see what happens when AI does your thinking. The question here is narrower: if you keep a “second brain”, what should the assistant be allowed to touch?

Does an AI assistant replace a second brain?

The idea of an external memory is older than any app. In the July 1945 issue of The Atlantic Monthly, Vannevar Bush described the memex, a desk at which a person would store books, records and communications and consult them at speed, “an enlarged intimate supplement to his memory”. Its essential feature was associative indexing: “The process of tying two items together is the important thing.”

A well-documented working example came from a sociologist. Bielefeld University’s project on Niklas Luhmann’s papers describes his card file as roughly 90,000 notes written between 1951 and 1996, a collection it calls his intellectual autobiography and the machine he relied on to develop his theory and publications; his output ran to nearly 600 publications, including more than 40 monographs. That is an average of about 2,000 cards a year, or roughly 5.5 a day, over 45 years (CEOtudent calculation). The archive was a by-product of daily thinking, not a substitute for it. Today the phrase is associated with Tiago Forte’s 2022 book Building a Second Brain, whose method is not reproduced here.

What chat assistants changed is that the archive can now talk back. They already act as a search layer: in OpenAI’s own classification of a representative sample of consumer ChatGPT conversations, “Seeking Information” grew from 14% to 24% of usage between July 2024 and July 2025 (NBER working paper with OpenAI-affiliated authors, not peer-reviewed).

So an assistant can replace much of the storage and retrieval work of a second brain, but not the processing that happens in your head while you build it. Psychologists call using tools to reduce the mental demands of a task “cognitive offloading” (Risko and Gilbert, 2016). A notebook offloads remembering; an assistant can also offload summarising, linking and drafting. The research shows which of those handoffs cost you something.

What does the research say each notes operation does to your memory?

Table 1 joins lines of research that are rarely read together: one row per operation a notes system performs, with the design, the result as printed and the implication for AI.

Table 1. What each notes-system operation does to your own memory (CEOtudent analysis of Bertsch et al. 2007, McCurdy et al. 2020, Storm and Stone 2015, Grinschgl et al. 2021, Camerer et al. 2018, Rowland 2014, Kornell et al. 2009 and Fan et al. 2025)

Notes operation	Study and design	Result as printed	Implication for AI handoff
Writing it yourself (generate vs read)	Bertsch et al. 2007: 445 effect sizes, 86 studies (abstract only); McCurdy et al. 2020: 310 experiments	Effect size .40; mean difference of .102 in proportion remembered	An AI summary you only read is the “read” condition
Saving to an archive before new learning	Storm and Stone 2015, three experiments (abstract only)	Saving one file improved memory for the next; no benefit when saving was “deemed unreliable”	Storage is safe to offload if the archive is trusted
Offloading while doing the task	Grinschgl et al. 2021, three experiments, N = 172 each	Performance up, memory down; correlations of -.80 and -.55 between lookups and later memory	Decide in advance what to remember, then look it up less
Building maps vs recalling	Karpicke and Blunt 2011, as described in Camerer et al. 2018	One week later: 0.67 retrieval practice vs 0.45 concept mapping (N = 40); replicated (N = 49)	AI may propose links; recall stays yours
Self-testing during review	Rowland 2014: 159 effect sizes, 61 studies	g = 0.50; 0.73 with feedback vs 0.39 without	AI writes questions and feedback; you answer
Attempting before seeing the answer	Kornell, Hays and Bjork 2009, Experiment 1	Recall .41 after failed attempts vs .31 read-only, d = 0.58	Try first, then prompt
Polishing output with an assistant	Fan et al. 2025, randomised lab study, 117 students	Essay scores improved more with ChatGPT; knowledge gain and transfer “not significantly different”	A better draft is not evidence of learning

Writing it yourself: the generation effect

Slamecka and Graf (1978) ran five experiments comparing words participants generated from a cue and a rule with the same words simply read; in all cases the generate condition was superior. The effect held across recognition, free and cued recall, appeared with only 12 participants, and did not depend on whether people were told a test was coming.

Two meta-analyses agree on direction. Bertsch and colleagues (2007) summarised 445 effect sizes across 86 studies and found an average of .40, “almost half a standard deviation” (abstract only; the full text sits behind a bot check). McCurdy and colleagues (2020) analysed 126 articles, 310 experiments and 1,653 estimates and put the advantage at .102 in the proportion remembered, about 10 percentage points. Larger samples showed smaller effects; extrapolated to an infinite sample the advantage fell to .080, about a fifth smaller (22%, CEOtudent calculation), but stayed significant. Lower-constraint generation, with fewer rules on what to produce, beat medium- and higher-constraint generation, and retention intervals over a day added approximately 7 percentage points.

For notes, the translation is direct: an AI summary you read is the read condition; a note written from memory, in an open format rather than a template, is the generate condition.

Saving to a trusted archive: offloading that helps

Not all offloading hurts. In three experiments by Storm and Stone (2015), saving one file before studying a new one significantly improved memory for the new file. The benefit disappeared when saving was deemed unreliable and when the saved file was too slight to interfere with the next (abstract only). Fellers and Storm (2024) found the effect when only some items were saved, by button or by hand; Buchli and Storm (2025) found it larger in people with higher working memory capacity (both abstracts only). Putting things away somewhere you trust frees attention for what comes next, and any tool, including an assistant, can do that job.

Offloading while you work: performance up, memory down

Grinschgl, Papenmeier and Meyerhoff (2021) ran three experiments with 172 participants each, 516 in total. Participants copied a pattern of objects from a model window they could reopen, and the researchers varied what each look cost and whether a memory test was announced. Offloading “accelerated task processing” but interfered with memory for the material. In Experiment 3, among people free to choose, the correlation between how often they reopened the model and how well they later remembered the objects was -.80 when they did not expect a memory test and -.55 when they did. People forced to offload maximally, but told a test was coming, almost completely counteracted the damage. Intention matters, so a second brain should record not only what is stored but what you intend to remember.

Linking and mapping versus recalling

Linking is central to second-brain methods, and Bush would agree. But building a map is not practising recall. In Karpicke and Blunt (2011), as described in the 2018 replication project’s pre-registered materials, participants who practised retrieval scored a mean of 0.67 on a test one week later, against 0.45 for those who built concept maps (N = 40): a difference of 0.22, or 49% more (CEOtudent calculation). The 2018 replication with 49 participants reproduced the result at 0.638 of the original effect size. Links make later retrieval possible, by machine and by you, but they are not a memory technique on their own. Suggesting links is a reasonable AI job; recall is a human one.

Testing yourself, especially after an attempt

Rowland (2014) pooled 159 effect sizes from 61 studies reported between 1975 and 2013 and found a testing effect of g = 0.50 (95% CI 0.42 to 0.58) against restudying. With feedback it was 0.73, against 0.39 without: a gap of 0.34, close to twice the size (CEOtudent calculation). Effects were larger at retention intervals of at least a day (0.69) than under a day (0.41). In the high-exposure subset (studies with feedback or initial test performance above 75%), recall-type initial tests (cued recall 0.72, free recall 0.81) beat recognition (0.36), an argument for open questions over multiple choice.

Kornell, Hays and Bjork (2009) justify an “attempt first” rule. In Experiment 1, participants faced fictional trivia questions they could not answer. Trying anyway before seeing the answer produced final recall of .41, against .31 when question and answer were shown together, d = 0.58: a 0.10 difference and a 32% relative gain (CEOtudent calculation).

Polishing output with an assistant

Fan and colleagues (2025) randomised 117 university students to an essay writing and revision task supported by ChatGPT, a human expert, writing analytics tools, or no extra tool. The ChatGPT group improved its essay scores more, but its knowledge gain and transfer were not significantly different, and the authors warn of “metacognitive laziness”. As writing as thinking argues, an assistant can raise the quality of the output without raising the quality of your understanding.

Did the “Google effect” survive replication?

The 2011 study by Sparrow, Liu and Wegner is often cited as evidence that search and storage change what people remember, and its experiments are summarised elsewhere on CEOtudent. The retest is the part that matters here. The Social Sciences Replication Project (Camerer and colleagues, 2018) replicated 21 experimental studies published in Nature and Science between 2010 and 2015, with samples on average about five times larger. It found a significant effect in the same direction for 13 (62%), with replication effect sizes on average about 50% of the originals. Three of the 21 bear directly on notes and memory practice.

Table 2. Memory studies in the 2018 Social Sciences Replication Project

Original study	Hypothesis tested	Original r (N)	Stage 1 r (N)	Pooled Stage 2 r (N)	Replicated?	Relative effect size
Sparrow, Liu and Wegner 2011	Study 1: computer terms more accessible after hard trivia questions (Stroop task)	0.368 (69)	0.110 (104)	0.050 (234)	No	0.299 at Stage 1; 0.135 pooled
Karpicke and Blunt 2011	Retrieval practice beats concept mapping on a test one week later	0.602 (40)	0.384 (49)	Not run	Yes, at Stage 1	0.638
Pyc and Rawson 2010	Test-restudy practice beats restudy for retrieving memory mediators	0.377 (36)	0.148 (132)	0.150 (306)	Yes, at Stage 2	0.394 at Stage 1; 0.398 pooled

Source: Camerer et al. (2018), Supplementary Information, Supplementary Tables 1, 3 and 4, figures as printed. Relative effect size is the replication r divided by the original r; CEOtudent recomputed each ratio. A second data collection (Stage 2) was run only when Stage 1 did not replicate.

Three details keep this honest. The project tested only Study 1 of the Sparrow paper, a reaction-time measure of whether computer-related words became more accessible; it says nothing about the paper’s other experiments. The team noted that the original paper reported a sample of 46 while 69 participants were actually used, and that the original authors did not respond to requests for materials, so the team programmed the task itself. The pooled replication used 234 participants, 3.4 times the original sample (CEOtudent calculation), and found an effect close to zero (r = 0.050, p = 0.449).

The two retrieval studies replicated, with smaller effects; the search-priming study did not. For a second brain, the better-supported worry is not that retrievable information rots memory by existing, but that skipping your own generation and recall, as in the Grinschgl and Fan results, leaves less in your head. More on the debate in is AI making you worse at thinking.

Which parts of a second brain must stay yours?

The evidence sorts notes operations into two kinds. Storage, sorting and search are offloading and, with a trusted archive, cost little for material you do not need to hold in your head. Generation, linking decisions and retrieval are processing; they pay off only if you do them. The framework turns that into an architecture.

Table 3. The Attempt-First Stack: four layers, one rule each (CEOtudent editorial framework)

Layer	What it holds	Who writes	What AI may read	The rule	Evidence reason
1. Capture inbox	Clippings, transcripts, saved files, raw ideas	You or AI	Everything in it	One trusted place; anything unprocessed after 7 days is archived or deleted	Saving helps only when reliable (Storm and Stone)
2. Own-words notes	Short notes written from memory, one idea each	You only	Yes, to search and quiz you	AI never rewrites this layer; every note carries a provenance label	Generate beats read (Bertsch; McCurdy)
3. Links and maps	Links, topic maps, open questions	You decide; AI may propose	Yes	A proposed link enters only with one sentence in your words on why	Mapping lost to recall in a replicated study (Karpicke and Blunt)
4. AI compute over the archive	Search results, clusters, drafts, review questions	AI drafts, you edit	Layers 1 to 3, inside your privacy boundary	Attempt first; output never enters layer 2 unless rewritten	Attempts and feedback build memory (Kornell; Rowland); polish is not learning (Fan)

Three standing rules make it work:

1. Provenance label. Every note says whether it is your words, a source quote or AI output. Without it, the archive fills with text you never processed and cannot tell apart.
2. The attempt-first gate. Before prompting on anything you want to understand, write your own answer, outline or guess, even a bad one.
3. No back-flow. AI output lives in layer 4 and moves into layer 2 only after you rewrite it from memory.

The prompts used in layer 4 belong in a prompt library, and the wider discipline of shaping what the model sees is context engineering.

Which tasks go to you, to AI, or to you-then-AI?

The decision table covers twelve common notes tasks, with an owner and the evidence behind each choice.

Table 4. Notes task routing (CEOtudent editorial framework)

Task	Owner	How	Evidence reason
Summarise a paper you need to understand	You-then-AI	Write five sentences from memory, then ask AI what you missed or got wrong	Generation plus feedback (McCurdy; Rowland)
Summarise a paper you only need to file	AI	Store the summary in layer 1, labelled as AI output	Trusted storage (Storm and Stone)
Find an old note	AI	Search across layers 1 to 3	Retrieval of stored items
Tag, file and de-duplicate the inbox	AI	Weekly pre-sort before your review	Storage work
Capture a meeting (raw transcript)	AI	Transcript goes to layer 1 only	Storage
Hold the decisions from that meeting in your head	You	Write decisions and owners from memory within a day	Intention to remember changes the cost of offloading (Grinschgl)
Write a permanent note in your own words	You	One idea, open format, no template	Lower-constraint generation remembered best (McCurdy)
Decide which notes are worth keeping	You	Weekly, during the review	Relevance is the owner’s judgment
Generate connections between notes	You-then-AI	List two or three links yourself, then ask AI for more; accept each with a reason	Attempt first (Kornell); mapping is not recall (Karpicke and Blunt)
Spaced review questions	You-then-AI	AI drafts open questions from your notes; you answer before seeing feedback	Testing effect, larger with feedback and after a day (Rowland)
Draft a memo from your notes	You-then-AI	Write the claim and outline first; AI drafts; you edit	Better drafts without more learning (Fan)
Search the archive before a decision	You-then-AI	Write your current position, then ask AI for notes that support and contradict it	Attempt first (Kornell)

The tally is 3 tasks for you alone, 4 for AI alone and 5 for you-then-AI (CEOtudent count). Very little must be done without AI, but almost everything meant to change your thinking needs you to go first. Capture norms for meetings are in how AI meeting notes changed the meeting, and Socratic prompting turns the assistant into a questioner for learning tasks.

How do you set the system up in one week?

The CEO move is a design decision, not a purchase: treat setup as writing an operating model rather than choosing software. This checklist works with any notes app and any assistant.

Week 1 setup checklist (CEOtudent editorial framework)

• Day 1: one inbox, one archive. Freeze duplicates; trust is a precondition of the saving benefit.
• Day 1: add a provenance field with three values: own words, source, AI output.
• Day 2: write the AI rules in your context file or project instructions: what the assistant may read, which layer it may write to, and the attempt-first rule.
• Day 2: draw your privacy boundary. List folders that never go to an outside tool (client material, other people’s personal data, anything your employer restricts) and check the tool’s retention and training settings.
• Day 3: seed layer 2 with 10 short own-words notes from recent work, written from memory.
• Day 4: save a question prompt that turns each own-words note into 2 open questions, never multiple choice.
• Day 5: book the weekly review in a fixed slot.

Table 5. The weekly 20-minute review (CEOtudent editorial framework)

Block	Minutes	What you do	Who does what
Inbox triage	5	Keep, file or delete what AI pre-sorted during the week	AI sorts, you decide
Cold recall	7	Answer last week’s open questions without looking, then check	AI asks and gives feedback, you answer
Promote and link	5	Rewrite one or two captured items as own-words notes; add one link with a reason	You write, AI may propose links
Next week’s remember list	3	Name three things you intend to hold in your head, not just in the archive	You

The blocks add to 20 minutes. Recall comes before promotion so the hardest block happens while attention is fresh.

What are the failure modes?

• The archive nobody reads. Capture without retrieval is storage, not learning. Skip the review for a month and the system becomes a filing cabinet; the fix is recall, not more capture.
• AI-written notes you never processed. Summaries made at capture look like progress. Without a provenance label they blend with your own words, and the generation step never happens.
• The untrusted archive. Several half-synced apps and no single source of truth. When saving is unreliable its benefit disappears, and you start holding things in your head out of doubt.
• Link inflation. Hundreds of AI-suggested links nobody can explain. A link without a written reason is decoration.
• Prompting before trying. Asking first feels efficient and skips the attempt that prepares memory for the answer.
• Feeding the whole archive to a vendor. Notes often hold other people’s information, client material and unpublished drafts. Before connecting an archive to an outside service, check what it retains, whether it trains on your data, who can access it, and what your employer’s policy says, and keep the boundary in writing. This is general information, not legal advice; your organisation’s rules and applicable law decide what is allowed.

What this does not mean

The evidence is strong on direction and weaker on transfer to real work. The studies cited here mostly use word lists, trivia questions, short texts or copying tasks, tested after minutes, days or a week, not months of professional reading. This review found no controlled trial of a full second-brain system with an AI layer, so Tables 3 to 5 are a design built from component findings, not a tested intervention.

Writing a summary is not a guaranteed memory technique. Reviewing ten learning techniques, Dunlosky and colleagues (2013) rated practice testing and distributed practice as high utility and summarisation, highlighting and rereading as low utility; summarising helped some students on some tasks, under limited conditions (abstract only). That is why the framework pairs own-words notes with recall instead of treating note-writing as the goal.

The failed Sparrow replication does not show that offloading is harmless; it shows that one priming result did not hold up. The Grinschgl experiments are better evidence that offloading during a task can cost memory, most clearly when you have not decided to remember. And effect sizes such as .40 or g = 0.50 are averages across very different studies; results vary with the material and how long you need to keep it.

Frequently asked questions

Does an AI assistant replace a second brain?
It replaces much of the storage and retrieval: finding notes, sorting an inbox, drafting from material already written. It cannot replace the generation, link decisions and recall that build your own memory, because those effects happen only when you do the work.

Should AI summarise reading notes?
For material you only need to file, yes, labelled as AI output. For material you need to understand, write a short summary from memory first, then ask AI what you missed. Generating beat reading across 86 studies in the Bertsch meta-analysis.

Is the “Google effect” real?
The priming result tested by the 2018 replication project was not reproduced: the pooled effect was 0.135 of the original. That test covered only Study 1 of the 2011 paper. Separate experiments, such as Grinschgl and colleagues (2021), do show that offloading during a task can reduce memory for the offloaded material.

What is the attempt-first rule?
Before prompting an assistant about anything you want to learn or decide, write your own answer, outline or guess. In Kornell and colleagues’ Experiment 1, even unsuccessful attempts raised later recall from .31 to .41.

How much time does the system take each week?
One 20-minute review: 5 minutes of inbox triage, 7 of cold recall, 5 of promoting and linking, and 3 to set next week’s remember list, after about five short setup sessions in the first week.

Is it safe to give an AI tool a whole notes archive?
It depends on what the archive contains and on the tool’s retention, training and access settings. Keep a written boundary for client material and other people’s personal data, follow your employer’s policy, and treat this as general information rather than legal advice.

Which notes app works best?
The research does not favour any app. It favours a single trusted archive, notes in your own words, and regular recall with feedback; any tool that supports those three habits will do.

Sources

1. Bertsch, S., Pesta, B. J., Wiscott, R. and McDaniel, M. A. (2007). The generation effect: A meta-analytic review. Memory and Cognition, 35(2), 201-210 (abstract only, PubMed record).
2. McCurdy, M. P., Viechtbauer, W., Sklenar, A. M., Frankenstein, A. N. and Leshikar, E. D. (2020). Theories of the generation effect and the impact of generation constraint: A meta-analytic review. Psychonomic Bulletin and Review, 27(6), 1139-1165 (publisher version, Maastricht University repository).
3. Slamecka, N. J. and Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 592-604.
4. Storm, B. C. and Stone, S. M. (2015). Saving-enhanced memory: The benefits of saving on the learning and remembering of new information. Psychological Science, 26(2), 182-188 (abstract only, PubMed record).
5. Fellers, C. and Storm, B. C. (2024). The saving enhanced memory effect can be observed when only a subset of items are saved. Memory and Cognition, 52(6), 1325-1337 (abstract only, PubMed record).
6. Buchli, D. R. and Storm, B. C. (2025). Working memory capacity and the saving-enhanced memory effect. Memory, 33(8), 1015-1030 (abstract only, PubMed record).
7. Grinschgl, S., Papenmeier, F. and Meyerhoff, H. S. (2021). Consequences of cognitive offloading: Boosting performance but diminishing memory. Quarterly Journal of Experimental Psychology, 74(9), 1477-1496 (PubMed Central full text).
8. Rowland, C. A. (2014). The effect of testing versus restudy on retention: A meta-analytic review of the testing effect. Psychological Bulletin, 140(6), 1432-1463 (author-uploaded copy).
9. Kornell, N., Hays, M. J. and Bjork, R. A. (2009). Unsuccessful retrieval attempts enhance subsequent learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 35(4), 989-998 (author copy).
10. Camerer, C. F., Dreber, A., Holzmeister, F. and colleagues (2018). Evaluating the replicability of social science experiments in Nature and Science between 2010 and 2015. Nature Human Behaviour, 2(9), 637-644, abstract and Supplementary Information (source of the Sparrow, Liu and Wegner 2011, Karpicke and Blunt 2011 and Pyc and Rawson 2010 figures; those original papers were not opened).
11. Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J. and Willingham, D. T. (2013). Improving students’ learning with effective learning techniques. Psychological Science in the Public Interest, 14(1), 4-58 (abstract only, PubMed record).
12. Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X. and Gasevic, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489-530 (figures from the arXiv 2412.09315 version).
13. Risko, E. F. and Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676-688 (abstract only, PubMed record).
14. Chatterji, A., Cunningham, T., Deming, D. J., Hitzig, Z., Ong, C., Shan, C. Y. and Wadman, K. (2025). How People Use ChatGPT. NBER Working Paper 34255 (not peer-reviewed; several authors affiliated with OpenAI).
15. Bush, V. (1945). As We May Think. The Atlantic Monthly, July 1945 (electronic reproduction).
16. Bielefeld University, Faculty of Sociology. Niklas-Luhmann-Archiv, research project “Niklas Luhmann – Theorie als Passion”, project description (German).
17. Forte, T. (2022). Building a Second Brain. Atria Books (named only; catalogue record checked, method not reproduced).

Tables 1, 3, 4 and 5 are CEOtudent analyses or CEOtudent editorial frameworks; Table 2 reproduces figures as printed in Camerer et al. (2018), with the ratios recomputed by CEOtudent. Derived figures (the Luhmann averages, the 516 total, the 22% shrinkage, the 0.34 gap, the 32% and 49% relative gains, the 3.4 sample ratio and the routing tally) are CEOtudent calculations. Deliberately not used: the original Sparrow, Liu and Wegner effect sizes (covered in other CEOtudent posts), the Adesope, Trevisan and Sundararajan (2017) practice-testing meta-analysis (no open copy found), and forgetting-curve percentages.

---

This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.

This post is also available in: