lifelogger

Problem

I generate a lot of small signals during the day — git commits, browser history, calendar events, terminal sessions, music. None of those individually mean much. Together they're a useful proxy for "what did I actually spend the day on?" but no single tool aggregates them in a way that's interpretable a week later.

Solution

A daemon that quietly ingests activity from multiple sources, batches it, and runs LLM summarization on a configurable cadence (hourly digest, daily wrap, weekly arc).

How

• Stack: Python.
• Sources: git, browser, calendar, terminal, foreground app — extensible.
• Storage: local-first; LLM calls are batched to keep token cost bounded.

The design treats every signal as an event with a timestamp, source, confidence, and privacy level. Raw events are cheap to collect but too noisy to read directly, so the important layer is the reducer: it groups nearby events into work sessions, removes obvious duplicate evidence, and produces a compact packet that an LLM can summarize without seeing the entire day.

The summarizer is intentionally staged. Hourly summaries answer "what just happened?" Daily summaries answer "what did this day become?" Weekly summaries look for continuity, stalled threads, recurring distractions, and projects that quietly consumed more time than expected. That separation keeps each prompt small and lets later summaries cite earlier ones instead of replaying raw logs.

The privacy boundary is local-first. Sensitive sources can stay as counts or labels, and the daemon can redact URLs, contact names, or terminal content before any model call. The goal is not surveillance for its own sake; it is a memory prosthetic that leaves a useful audit trail while keeping raw behavior under local control.

Results

Active personal infra. The useful output is not a dashboard of every event; it is a small set of review surfaces: a daily narrative, a weekly arc, and selected updates that can feed back into this site.

Lessons

Batching is the difference between "neat idea" and "I can actually afford to leave it on." Without batching, the LLM cost dominates the value.