TL;DR: Use auto-logging instructions to turn Claude Code from a chat tool you use into a personal AI agent that learns and remembers automatically so that every conversation compounds.
I've been running my personal life through Claude Code for a few weeks now, and one instruction is key to how the whole system works: auto-logging. Instead of Claude asking what to track and where to put it, it tracks everything automatically and adds notes to right files.
This also enables the system to "self-improving": every conversation adds context, every person I mention gets remembered, and each session builds on the last.
The instruction
## AUTO-LOGGING BEHAVIOR **Do NOT ask before logging. Automatically:** 1. When user mentions a task → Add to appropriate file's TODO section 2. When user completes something → Mark done, move to daily log 3. When user shares notes/context → Add to daily log under today 4. When user shares product ideas → Capture in detail to notes/ 5. Every session → Add entry to log.md with summary
Once this was added this to my CLAUDE.md, every conversation accumulates personalized memory in the system. I mention a task and it's captured, say something's done and it's marked complete, share context and it's logged.
It turns out the experience is pretty smooth: I don't feel like I'm maintaining anything; I'm just having conversations, and the system grows from them.
Example uses
Me: "I need to follow up with Sarah about the proposal."
Claude: Adds to work.md:
- [ ] [P1] Follow up with Sarah about proposal @person:Sarah @followup
"Got it. Anything else on your mind?"
The task is now searchable forever - the files are going to grow over time, but the models can handle huge codebases, and so also huge md files! Next week, when I mention Sarah again, Claude already has context or can do a quick search to pull it up. A month from now, I can grep my entire history with her, the tagging system was designed by Claude where it self-documented it.
Git as memory
Since Claude Code already knows git, my repo has naturally become a long-term memory for the agent with full version history:
# See what changed git log --oneline -20 # Roll back if something breaks git checkout HEAD~1 -- work.md
Every commit is a checkpoint, so if Claude logs something wrong or you want to undo changes, you just roll back. There's no black box in how memory is stored, no trust required.
And markdown files scale. A tracking file can grow to thousands of lines and Claude handles it the same way it handles large codebases - just let it grow. In any other system you'd be pruning and archiving, but here you don't.
I also get the benefit of portability, where I can move to a different agent model in the future since I own all the files.
Starting out
I've been showing my setup to a few folks, and getting questions on how this all started. Turns out, it was a pretty small experiment to begin with, ~beginning 2026 where I was thinking through goals for the year. I started with three files:
personal/ ├── CLAUDE.md # System instructions (auto-logging lives here) ├── work.md # Work todos, daily log ├── personal.md # Fitness, health, goals ├── home.md # Household tasks, family items └── log.md # Session activity log
None of the file are hand-written, Claude does all the writing, while I usually dictate to it or give it instructions. Within the last few weeks of regular use, the system has expanded:
personal/ ├── CLAUDE.md ├── work.md ├── personal.md ├── home.md ├── food.md # Meal tracking ├── log.md ├── people/ # Files per person (emerged from @person tags) │ ├── sarah.md │ └── michael.md ├── tasks/ # Folders for complex projects │ └── active-trip-planning/ ├── notes/ # Reference material and ideas └── claude-todos.md # Claude's own checklist
I didn't design this structure upfront, it emerged from conversation and occasionally asking Claude to reorganize to be more efficient. One thing I do really badly is remembering things about people, and so Claude helped me create a people folder to make notes. When I had a multi-session project, Claude proposed a task folder, now I've more than 30 tasks in different states in there covering everything from trip planning, backyard work, things I'm learning, gym workout routines, to work ideas.
Content routing
One important thing for auto-logging to work well is to have routing instructions in my CLAUDE.md, these were all setup automatically by Claude!
### Content Routing - Work tasks → work.md - Household, family, errands → home.md - Fitness, health, personal goals → personal.md - Info about people → people/ folder - Product ideas, strategy → notes/work-product-ideas.md - Complex multi-step projects → tasks/ folder If ambiguous, make best guess. I can correct.
Now when I ramble about multiple topics, each piece lands in the right place:
Me: "The X meeting went well. Also need to book a dentist appointment. And I had an idea for a new feature - what if we added..."
Claude: Updates three files simultaneously:
- Adds meeting notes to work.md daily log
- Adds dentist task to home.md TODOs
- Captures feature idea in detail to notes/
I didn't have to think about where anything goes - I just talked. One outcome of this is I've never felt as organized in my life as before!
Power of Markdown (text) files
Because everything is text files, it's all searchable. Claude has figured out that it shouldn't read entire files into memory, instead, it should be reading just either the top (using head) or bottom (using tail). If it needs more context, the agent can figure out how to get more data into it's context window.
# Find all entries tagged @run grep -n "@run" personal.md # Find all P0 (urgent) tasks across files grep -n "\[P0\]" work.md home.md personal.md # Find everything about a person grep -rn "Sarah" people/ *.md # Find all entries from a specific date grep -n "## 2026-01-15" *.md # Find all incomplete todos grep -n "\- \[ \]" *.md # Find items with deadlines grep -n "@deadline" *.md # Find all insights (tagged entries) grep -n "@insight" personal.md
No black box memory system
Most AI memory systems are opaque - you have no idea what the model "remembers" or how it uses that information. I've tried the ChatGPT memory system and it's kinda awkward to have this list of memories that I can delete and it's not easy to edit or create large memories.
Here, memory is text files I control. I can open them in any editor, read exactly what's stored, delete anything I don't want kept. Version controlled, so nothing disappears unexpectedly.
This transparency is why auto-logging feels safe. I'm not trusting an opaque system with my life data - I'm writing to files I own.
Gotcha: self-contained entries
I also learned to tell Claude to capture everything in text, not references. This came up after I started auto-syncing more data into Claude, particularly, Apple Health and Strava data. Sometimes Claude would make references to look at the app for more detail, while instead I wanted it to have full complete memory so that later we can do some analysis.
Claude came up with a prompt to make this work well!
### Self-Contained Logging Capture EVERYTHING in text so I never need to open another app. When logging activities (runs, meetings, trips): - DO NOT write "see Strava" or "see app" - DO describe routes, data, and details in full text - DO extract all metrics into the log Example - Running routes: - Bad: "Route: see Strava map" - Good: "Route: Started at Heritage Oaks Park → south on Springer → west on Loyola Dr → loop through Foothill area → return via residential"
The tracking system becomes the single source of truth - I can search and review everything without opening other apps. I've been tinkering with getting it to write scripts to analyze my data and it's actually pretty interesting.. for another post.
Getting started
If you want to try this out, create a brand new repo, create a CLAUDE.md file in a git repo and tell Claude the following!
# Life System Instructions ## Auto-Logging **Do NOT ask before logging. Automatically:** 1. When I mention a task → Add to appropriate TODO section 2. When I complete something → Mark done, move to daily log 3. When I share context → Add to daily log ## File Structure - work.md: Work tasks, meetings, daily log - personal.md: Health, fitness, personal goals - home.md: Household tasks, family ## Session Start 1. Read CLAUDE.md 2. Read work.md, personal.md, home.md TODOs 3. Surface urgent items and upcoming deadlines 4. Ask what I want to focus on
Create work.md, personal.md, and home.md with a simple structure:
# Work ## Active TODOs (empty - will populate from conversation) # Daily Log (reverse chronological entries go here)
Then just start talking. The system builds itself from there.
#agentifying-life #agents #claude-code #personal-productivity