Memories
Praxen Inc. keeps workspace memory as markdown files in a git repository, one repository per workspace. Agents read it and write to it, and so can you.
What it does
Memory records confirmed findings about your system: a failure mode, a recurring pattern, or an operational fact backed by evidence. When an investigation confirms a root cause, the agent writes it to a file, so the next agent starts from it and does not find it again. Agents do not store preferences, unverified claims, or configuration already visible in the topology.
How the repository is laid out
| Path | Written by | What it holds |
|---|---|---|
MEMORY.md | Praxen Inc., each night | The index of what is known about the workspace. |
daily/ | Praxen Inc., each night | One note for each day, with the patterns and signals of that day's threads. |
preferences/ | You | Files you edit to tell agents how your team works. |
| Other markdown files | Agents | One finding in each file, with the resource and the date in each entry. |
Each file starts with one line that says what it covers.
How memory reaches a thread
MEMORY.md and the three most recent daily notes appear in the briefing of every new thread. They show the state as of the nightly update. Agents clone the repository into the thread sandbox the first time the sandbox is used, and read it with grep or rg for anything the briefing does not show. Files saved after the nightly update are in the repository when the agent looks.
When a turn ends, Praxen Inc. commits and pushes what the agent changed. If two changes conflict, the agent opens the files that Praxen Inc. names, keeps every fact that is still true, and removes the conflict markers.
Edit memory yourself
Edit files under preferences/ to give agents standing guidance. Praxen Inc. rewrites MEMORY.md and writes the daily/ notes each night, so do not edit those. To record a finding from a thread, ask in plain words: "Save a note: checkout-api integration tests fail intermittently because the payment vendor's sandbox times out under parallel load, confirmed from pipeline logs showing 504 responses."