Attention Scarcity and Dunbar's Number in Personal Relationship Management

Neglected friendships usually drift without anyone deciding to drop them. The cause is an allocation problem: a finite budget of attention spread across a network larger than it can cover, with nothing in place to notice which ties are going without. The argument here is that personal relationship management should aim to prevent decay across a layered network under a hard limit on attention. Importing the CRM pattern of pipelines, stages and conversion measures the wrong thing, and it also changes what it measures. The sections below give a formal statement of the attention budget, a way to score a tie without turning a person into a lead, and design rules for measuring parts of private life without damaging them.

Dunbar's Layered Network and the Attention Budget It Implies

Dunbar's work on primate group size started as an argument about the limits of the brain. Neocortex volume predicts the size of group an animal can maintain, and for humans the extrapolation gives a figure of roughly 150 (Dunbar, 1992). The number itself has become a slogan. For design, the more useful finding is that the 150 are arranged in layers, each roughly three times the size of the one inside it: an innermost group of about five, then fifteen, then fifty, then the familiar hundred and fifty. Each layer needs a different contact frequency to stay where it is. In Dunbar's account of grooming and language, conversation is the human substitute for physical grooming, so the cost of maintaining a tie is paid in time (Dunbar, 1996).

Seen that way, this is the problem Simon described: in an information-rich environment, the scarce resource is the attention needed to process information (Simon, 1971). In a social network, the scarce resource is the number of meaningful contact events a person can manage in a year, however much affection or willingness there is. That gives a budget constraint you can write down directly:

maintenance budget:   Σ_i  n_i · f_i · c_i   ≤   A

n_i = members of layer i        f_i = contacts per year required by layer i
c_i = cost per contact (time)   A   = total attention available per year

Here is an illustrative example built on stated assumptions. Nothing in it was observed. Take the exclusive layer sizes implied by the standard nesting (5, then 10, then 35, then 100), give each layer the contact frequency it needs to keep from drifting outward, and assume a flat twenty minutes per contact event.

Under these assumptions the network costs about 620 deliberate contact events and roughly 207 hours a year, which is around four hours a week, every week, with no end date. The figure is not a measurement, but its order of magnitude is what matters here: maintaining a network at full size is a part-time commitment, and a tool that ignores the constraint hands the user a plan nobody could carry out. Storing 150 people is trivial, so the main job of such a system is allocation, deciding which of them get one of this week's few slots.

Relationship Decay Models and Intermittent Maintenance

The second structural fact is that ties decay without contact. With no maintenance, closeness drifts toward the network's outer layers, and a tie that was in the sympathy group two years ago is in the active network today without either person having chosen that. A simple model captures the shape: between contacts, tie strength follows s(t) = s₀ · e^(−t / τ), where τ is a per-relationship time constant, short for a colleague met at a conference and long for a sibling. Each contact pushes s back up, and the interval between contacts decides how far it falls in the meantime. It is the same shape as in content rotation against the forgetting curve, where a decaying quantity is restored by spaced re-exposure instead of constant presence.

Two design implications follow, and both run against what a task manager would suggest. First, the signal that matters is the time since last contact measured against that relationship's own τ. Three months of silence is worrying for a close friend and normal for a former manager, so a single global "you haven't spoken in 90 days" threshold raises false alarms for some ties and misses real decay in others. Second, maintenance happens in bursts. A two-hour conversation once a quarter can hold a tie that fifteen daily messages would not, because the size of the reset matters more than the number of messages. A system built on streaks or daily engagement brings the retention mechanics of consumer apps into a domain where nothing justifies them.

Why the Sales-CRM Objective Function Fails in Private Life

Personal relationship tools are often described as "a CRM for your life", and that analogy is where the design goes wrong. A sales CRM optimises a well-defined objective: move a finite set of opportunities through ordered stages toward a conversion event, subject to quota, within a period. In private life almost every element of that objective is either missing or harmful.

There is no conversion event. A friendship has no closed-won state, the relationship is the process, and a stage model is a category error. There is no quota and no period either, so forecasting and pipeline weighting have nothing to work on. Nor is there a direction to push in, since the goal is to stop ties from decaying by accident when their current layer suits both people. Some relationships should stay at quarterly contact for good, and a system that treats that as a stalled deal only adds pressure.

Carrying the CRM over also brings a specific measurement risk. A pipeline view turns people into interchangeable units in a funnel, and a score borrowed from lead qualification implies that some contacts are not worth the call. Sales tooling is built to make that judgement. A personal system that makes it is doing something the user never asked for and would reject if it were stated plainly. Putnam's account of declining civic and social participation (Putnam, 2000) is relevant here. The deficit he describes is one of participation and time, and treating acquaintances as a managed portfolio does not obviously fix it.

Weak Ties and the Informational Value of the Neglected Layer

The layer a naive allocation is most likely to neglect is also the one with the highest informational return. Granovetter argued that new information, such as job openings, opportunities and introductions, arrives disproportionately through weak ties, because strong ties sit in the same dense cluster as you and know what you already know (Granovetter, 1973). Burt's structural-holes account puts the mechanism in terms of position: people who bridge otherwise disconnected clusters gain an advantage (Burt, 1992). Milgram's small-world experiments showed that short paths through such bridges exist in the first place (Milgram, 1967).

This is awkward for allocation. Weak ties, the fifty and the hundred and fifty rather than the five, are the ones with the lowest emotional salience and the largest τ-relative decay when neglected. A tool that only surfaces the relationships the user already thinks about every day adds nothing, because those ties maintain themselves. Where a tool can help is in surfacing the outer layers before they lapse, at a pace the budget can absorb. In our own work on RelationCRM, this is why the system scores 150+ relationships instead of a curated shortlist, since the people on a shortlist are the ones the user does not need help with.

Scoring a Tie Without Reducing a Person to a Lead

If ranking needs a score, the design question is what the score must not flatten. A lead score has one dimension by design, the probability of revenue, and that is exactly what to avoid. RelationCRM scores relationships on five dimensions, so the reason a tie gets flagged stays readable: recency against that relationship's own baseline, reciprocity, depth, layer fit, and outstanding obligations. Two ties can have the same total and need completely different responses, and the breakdown shows that.

Three inputs do useful work here without much modelling. Importing WhatsApp history gives each relationship a baseline cadence, so τ comes from data and is not just assumed. Sentiment analysis over that history picks up the direction a relationship is moving in. That matters because a tone that is cooling while contact frequency stays the same is invisible to a system that only counts contacts. Extracted promises, the "let's do this in September" that both people meant, turn a vague obligation into a specific item. Reminders that come before a birthday, and not on the morning of it, follow the same idea, because a same-day notification produces a generic message. Message drafting in six tones and ten languages is there to bring the effort of reaching out down to something the budget can afford. As with any multilingual product, the tone work is closer to cultural rewriting than to translation.

The explicit non-goals matter as much as the dimensions. The score must never rank people by worth, never appear as a leaderboard and never be exported or shared. It must also never be presented as a judgement about a person, only about a pattern of contact. It is there to help the user allocate scarce attention, and the way it is shown should make that clear.

The Instrumentation Paradox in Measuring Relationships

Measuring a friendship can change it. Goffman's analysis of self-presentation treats social life as a performance with a front and a back region (Goffman, 1959). Instrumentation risks moving private relationships onto the front stage, where every contact is logged and scored. Once a message counts toward a metric, the sender can no longer be sure why they sent it, and neither can the recipient. The user may end up tending the record instead of the relationship.

We took that seriously, and it shaped several decisions. Scores never appear while the user is writing a message, so nobody is writing to a number. Nothing is ever visible to the other person, which keeps the tool entirely on one side. Contact is not gamified in any way: there are no streaks, completion percentages or social scores, because those mechanics turn intrinsic motivation into compliance. A reminder is framed as a prompt to make a decision, not a task to clear, and dismissing it is a perfectly good outcome. The system should be able to say that a tie is fading and that letting it fade is a legitimate choice.

Privacy as an Architectural Precondition for On-Device Relationship Data

Everything above depends on the most sensitive data most people have: their private conversations and an assessment of their closest relationships. That makes privacy an architectural precondition for the whole product. RelationCRM keeps relationship data in on-device storage, anonymises contacts to identifiers such as "Contact-001" before any model call, and runs a zero-retention policy on the inference path. Under the GDPR (Regulation (EU) 2016/679), data minimisation and purpose limitation are legal obligations, and processing conversation history makes them very concrete. The other people in an imported chat are data subjects who never installed anything. Keeping the data on the device and pseudonymising the model inputs is what makes that third-party problem manageable.

Limitations

None of this shows that the approach works. Dunbar's layer sizes are regularities across populations with wide individual variation, and 5/15/50/150 should be read as a central tendency, not as a specification for any one person. The exponential decay model is a convenient shape that has not been validated as a law, and we have not estimated τ against any outcome. The budget table is an illustration and does not measure anyone's network.

Most importantly, we have no evidence that a scored, reminder-driven system improves relationship outcomes. RelationCRM is in development, so this article explains the reasoning behind its design and has no results to report. The measures listed against the instrumentation paradox are hypotheses about avoiding a harm we cannot yet measure. It is still possible that the right conclusion is to leave some parts of private life unmeasured.

References

• Dunbar, R. (1992). Neocortex Size as a Constraint on Group Size in Primates.
• Dunbar, R. (1996). Grooming, Gossip and the Evolution of Language.
• Granovetter, M. (1973). The Strength of Weak Ties.
• Simon, H. A. (1971). Designing Organizations for an Information-Rich World.
• Burt, R. S. (1992). Structural Holes: The Social Structure of Competition.
• Milgram, S. (1967). The Small-World Problem.
• Putnam, R. D. (2000). Bowling Alone: The Collapse and Revival of American Community.
• Goffman, E. (1959). The Presentation of Self in Everyday Life.
• European Union (2016). Regulation (EU) 2016/679 (General Data Protection Regulation).