Multi-agent systems fail in a predictable way: agents either forget what the others did, or remember far too much of it.
Two ways memory goes wrong
Give every agent the full history and they drown in it. Context fills with other agents' half-finished thinking, costs climb and answers drift. Give each agent only its own notes and the handoffs break: the analyst re-reads the ledger the researcher already summarised.
Memory scoped to the job
Ordin keeps one memory per job, split into what is shared and what is private. Findings, decisions and outputs are shared. Scratch work stays with the agent that wrote it.
- Shared: facts, decisions, files and finished outputs
- Private: drafts, intermediate reasoning and retries
- Expiring: anything tied to a single step is dropped when the step settles
Handoffs stopped feeling like a game of telephone. Each agent starts exactly where the last one finished.
Engineering manager, logistics customer
Every write is replayable
Memory writes are part of the trace. When you replay a run, you see what each agent knew at each step, which makes the question of why it decided something answerable for the first time.
- Memory
- Multi-agent