Every agent Wayari puts on a job starts a fresh session. Nobody carries a whole day of chat into the next piece of work.
A fresh session has a cost. Each agent has to rediscover the command that runs the tests, and the check that broke last time and what fixed it.
None of that is in the code. Somebody knew all of it last Tuesday.
01a wayspace keeps a ledger
A wayspace is one repository's memory. Every agent that works on that repository can read it.
It holds a ledger. Each entry comes from something that already happened, carries the commit it happened at, and names the shift that learned it.
You can delete any entry in one line. Nothing in it is prose a model wrote about your code.
Anything that can be worked out from the repository belongs to the atlas instead, a map derived from git. A derived fact cannot go stale in the first place.
02three kinds of entry
The ledger holds three kinds of entry, and the list is closed:
- A command that passed a gate, with the commit it passed at.
- A check that failed, with its output and the diff that turned it green.
- A convention in your repository that a tool would not guess.
Wayari writes the first kind itself. It is standing where the check passed, so npm test in the ledger is a fact rather than a guess.
A command that failed is never recorded as a command. Nobody knows it works.
Each entry names the paths it depends on. When those paths change, the entry is stale, and Wayari drops it rather than serve it as current.
03what stays out
A summary of what a file does stays out, and so does an opinion about the architecture.
Your own CLAUDE.md and AGENTS.md are read where they are and never copied.
The reason is a study from ETH Zurich and LogicStar, from February 2026. Context files written by a model lowered agents' resolve rates and cost more to run.
Repository overviews did not help at all. Naming a repository's own tool moved an agent's use of it from 0.05 calls to 2.5.
Agents act on commands and ignore prose, so the ledger holds commands.
04the thread, read from anywhere
wayari thread is what people usually mean when they ask for a wiki. It is one record per repository of what happened, and any client can read it.
Nothing in it is written directly. Wayari folds it on every read from records that already exist, such as a channel's notices and a gate's verdict.
So the thread cannot go stale the way a wiki page does. There is no second copy of a fact waiting to disagree with the first.
The one thing it stores is a note, because a note is somebody's own words. wayari note writes one, and every note an agent writes carries its author.
Standing orders sit beside it. wayari standing add keeps what you would say to every agent before every job, and every shift receives it.
05an answer instead of an argument
Nobody in this category has published evidence that a context store makes agents better. Wayari has not measured it yet either.
What it has is a control arm. The atlas withholds its excerpt from one shift in five, and every ledger entry rides the same arm.
So whether memory makes a shift better will be a number rather than an argument. The number does not have to be flattering to be worth having.
