Every run starts warm. No agent begins from zero.
Your architecture, your conventions, and every hard-earned lesson your team has already learned travel with every agent, on every run. The longer Fleet Work runs, the sharper its judgment gets.
Three layers of memory, working together
Global memory carries the conventions that apply org-wide. Repository memory holds the patterns specific to each codebase. Session memory logs exactly what happened on a given run: files touched, decisions made, contracts uncovered. Each layer surfaces to the agent exactly when it's needed.
Global Memory
Org-wide conventions
Repository Memory
Per-codebase patterns
Session Memory
What happened, per run
Living documents, not a scrollback log
Instead of piling up a timeline, Fleet Work distills what happened after every story into tidy, deduplicated knowledge, patterns, decisions, gotchas, API contracts, organized by topic and folded straight into your global and repo memory.
- · Auth: token refresh must precede retry
- · Payments API: idempotency key required
- · Schema: soft-delete pattern for orders
Ask a question, get the decision that answered it
Past runs surface by concept rather than exact wording, so the agent walks in already knowing your setup. It follows the conventions your team already settled on instead of reinventing them, and sidesteps gotchas it's already hit once.
Give it a head start before day one
Load repository memory with your architecture and patterns up front, so the very first run reads like the hundredth. Memory is what separates an agent that guesses from one that knows, and it's fully yours: editable, inspectable, and always growing as more stories ship.