HQ: a memory system for agent work
Markdown, git, and a few conventions that let AI agents pick up work across sessions without re-explaining anything.
The biggest waste in working with AI agents is re-explaining context every session, and losing decisions to the scroll. The model is rarely the bottleneck; the memory between sessions is. HQ is my fix: a plain-files headquarters where every project has a channel, every day has a journal, and anything important gets written to disk the moment it's decided. It replaced a graveyard of abandoned Notion pages, half-finished wikis, and "I'll remember that" with one git repo that both I and every agent I work with treat as the source of truth.
Files over frameworks
No orchestration platform, no vector database, no dashboard. A git repo of markdown: an index you navigate like a book, per-project channel files where sessions leave handoffs for the next one, and most-important-task lists at daily through quarterly horizons. Any agent, or any human, cold-starts by reading four small files. The whole thing is greppable, diffable, and readable on a phone from the GitHub app, which is more than I can say for any knowledge tool I've paid for.
Write-through by default
The rule that makes it work: decisions and answers get written to disk verbatim as they happen, never batched for later. Context windows collapse, sessions get compacted, and "we discussed this an hour ago" is a lie agents tell themselves. Files don't. Session transcripts land in the day's journal folder automatically via hooks, and a review pass promotes anything worth keeping into the permanent structure. The system assumes forgetting is constant and designs for it, instead of hoping it won't happen.
Channels instead of chat history
Each project has a CHANNEL.md that works like a slow, permanent Slack channel: dated entries, newest context on top, decisions marked as decisions. When a fresh session starts on a project, it reads the channel and knows the state of play in thirty seconds, including things I'd personally forgotten. The channel outlives every individual conversation, which quietly fixes the worst property of chat-based work: the best thinking dies in the transcript.
The counter-example
This is the boring cousin of the agent-orchestration setups I wrote about in "The agent orchestration trap." No swarm, no CEO agent, no productivity theater, just a system of record that made one human plus a few agents genuinely faster: it carried a security teardown, this website, and a conference workshop through in the same week, with every decision traceable to a dated line in a file. The lesson I keep re-learning is that better memory does more for agents than more autonomy, and memory turns out to be a folder of markdown.