Andreas Heige

Project 1 of 3 · Daily, 08:00

Autograph

Coding-agent sessions and git history, written into a linked Obsidian vault every morning — by a model that never leaves the machine.

The idea

Claude Code, Codex, Copilot CLI and Pi keep separate session transcripts. Autograph collects them, plus git and the Claude Desktop chats I choose to save, and every morning at 08:00 writes the last seven days into one Obsidian vault: a note per day, a hub per project, a page per agent, linked to each other and to the commits.

The writing is done by a local model through Ollama, Gemma by default. Session records stay on the Mac; only the vault's Markdown goes to its git remote.

I had something like it on my old computer and wanted to rebuild it from scratch. The idea is a diary of what happens during long sessions, when agents run in auto mode. It also helps with context switching: when I have ten things going at once and have to pick them up tomorrow, it is what my popcorn brain needs.

An index first

The vault is not one pile. Autograph's raw capture of events sits in its own folders and is never read back as context. What agents do read is curated memory: one fact per file, each with a one-line entry in an index. Claude, Codex and Copilot all read the same index, through their own global instructions, so a lesson learned in one tool is there for the next.

The index is also the search. An agent skimming one line per fact is already doing the ranking, so there is no retrieval tool, and none is planned until a memory folder passes about a hundred files. A separate scheduled job builds a wiki from plans, curated memory and repository docs. At the time of the deep dive, its roughly 300 files did not justify a database or a memory search tool.

Machines propose; I decide. Nothing becomes a tag until I tick it in vocabulary.md. Memory works differently. A job called dreaming runs at 07:00: it extracts user and assistant dialogue from recent transcripts and proposes memory facts with evidence quotes. Opus reviews them in a fresh context and ticks the ones worth keeping; the next run applies them. The memory rules say to rewrite a confirmed correction in the existing file, keeping a "Previously" line. Unclear contradictions stay disputed until I resolve them.

The struggle

One limit was the size of what the model sees. Each unit of work goes to the model with at most 6,000 characters of evidence, and the cut was silent: it kept a chronological prefix and dropped the rest. A whole session used to be one unit, so a 12-hour session of 1,800 events reached the model as roughly its first 40 minutes. Everything after that was kept on disk and never made it into a note. Now sessions are split at idle gaps over an hour or when the evidence would overflow. The pipeline keeps up to five chunks per session and twenty per day, ranked by message counts. That limits how much one session can occupy; some chunks still stay out of the note.

The local model had its own habit. Gemma on Ollama spent its output budget on reasoning, so unattended jobs crawled or timed out: a wiki lint took 63 minutes, with 10 of 40 checks hitting the five-minute limit. With thinking off and answers capped, the lint ran in under two. The resulting memory fact is marked as added by dreaming.

Small bugs leave large holes in a graph. A newline carried over from git log landed inside commit links and broke 59 of them across five daily notes before one .strip() fixed it.

A daily note is the wrong shape for a long investigation anyway. When I want to go deep into something, it gets its own deep dive, one topic per file, next to the daily notes rather than inside them.

I finally solved context switching. The time it freed up went to sitting in two video calls at once; I suspect the answer is not a third.

  1. CollectSeven days of Claude Code, Codex, Copilot and Pi sessions, plus commits.
  2. GroupInto units of work per project and session, cut at idle gaps.
  3. Write, locallyOllama writes what was done, what went well, what was learned.
  4. Tag and linkOnly with terms you ticked; up to three related wiki pages.
Fig. 1 — A week of work, drawn into the vault at 08:00.
Reads
Four agents and git
Writes
Daily notes, projects, wiki
Model
Ollama, on the Mac
You decide
Which tags exist
Machines propose. You decide. Nothing becomes a tag until you tick it.