Quantifying Friendship: Analyzing Two Decades of Chat History

The concept of the "digital footprint" is often discussed in terms of privacy and surveillance, but rarely do we consider it as a mirror for self-reflection. When a user (valzevul) decided to analyze 20 years of their own chat history to answer the question "Am I a bad friend?", they transformed a mountain of unstructured text into a sociological map of their own life.

This experiment highlights a growing trend: using Large Language Models (LLMs) not just for productivity, but as a tool for personal archaeology—digging through decades of logs to find patterns that are invisible in the moment.

The Methodology of Personal Archaeology

Analyzing two decades of communication is a daunting task. The data is often fragmented across multiple platforms—IRC, MSN Messenger, ICQ, Facebook Messenger, and WhatsApp—each with its own export format and noise level.

To make sense of this data, the author leveraged LLMs to categorize contacts and analyze the nature of interactions. This approach solves a problem that previously required manual sifting or complex Natural Language Processing (NLP) pipelines. As one community member noted, LLMs are uniquely suited for this because they can "psychopathically chug along without flinching or cringing" through the embarrassing remnants of one's youth.

Technical Considerations for Chat Analysis

For those looking to replicate this analysis, several technical challenges and opportunities emerge from the discussion:

Noise Reduction: While manual filtering is one option, some suggest using TF-IDF (Term Frequency-Inverse Document Frequency). This method identifies words common across the entire corpus as noise, while highlighting words that appear frequently in a specific chat as significant markers of that relationship.
Entity Resolution: One of the hardest parts of long-term data is the shifting identity of contacts (e.g., nicknames on IRC vs. real names on Facebook). LLMs have proven surprisingly effective at stringing together these disparate identities by analyzing the context of conversations.
Platform Bias: It is crucial to remember that data is a proxy for behavior, not the behavior itself. A drop in "question rate" in a chat log might not indicate a decline in curiosity or friendship, but rather a shift in medium—such as moving from text to voice or video calls.

The Sociology of the "Friend Signature"

One of the most compelling aspects of the analysis is the attempt to quantify the "friendship" level based on communication patterns. The author categorized their social circle into close friends, regular contacts, and active acquaintances.

However, this quantification sparks a philosophical debate. Some argue that the ability to maintain meaningful contact with hundreds of people is an enviable skill, while others find the idea of a "personal CRM" (like MonicaHQ) to be an antithesis to organic human connection.

"I fear about getting too aware from the insights of the analysis, I fear my opinions my faith might change in certain people."

This sentiment highlights a primary risk of data-driven friendship: the danger of reducing a complex, emotional bond to a set of metrics. When we quantify our friends, we risk seeing them as data points rather than people.

The Digital Divide: Archivers vs. Erasers

The discussion reveals a stark divide in how people handle their digital history. On one side are the Archivers, who keep every hard drive since 2001 and view their logs as a valuable historical record. On the other side are the Erasers, who use disappearing messages on Signal or set their apps to auto-delete history every six months.

For the Erasers, the possibility of this kind of analysis is non-existent. For the Archivers, the risk is that their private thoughts from a decade ago are now "safe from being harvested by AI," potentially leading to a form of "mental cloning" if those logs were used to train a personalized LLM.

Conclusion

Analyzing 20 years of chats is more than a technical exercise; it is an act of introspection. While the data can provide objective insights into who we spent our time with and how we communicated, the true value lies in the questions it forces us to ask about our social habits. Whether we use LLMs to map our friendships or delete our history to live in the present, the digital record remains a powerful, if sometimes uncomfortable, testament to who we were and who we have become.