Why it matters: Multi-agent Claude pipelines drift and duplicate work not because the model forgets, but because nobody built a place for it to remember in the first place.
The Details:
A language model holds no state between runs. Its context window is working memory that vanishes when the session ends, so any decision that isn't written to a file the next agent is forced to open simply never happened as far as that agent is concerned. Treat this as an ordinary state-management problem: decide where facts live, who writes them, who must read them before acting.
Stale memory is worse than no memory. A file that was once true and silently drifted invites confident, wrong action, while an empty file at least sends the agent to check. Every entry needs an owner, a date, and a reason, so agents can tell a claim about the past from a guarantee about the present, and someone has to prune aggressively or the document rots into noise nobody trusts.
Split memory by function instead of dumping everything into one file. Durable rules and conventions belong in something like a project charter file; live status belongs in a short working-state document written as a contract, not a diary; and reasoning history, the part teams skip most, stops agents from rediscovering dead ends the system already walked. Structured handoff artifacts beat prose summaries because they define exact inputs and outputs instead of one agent's lossy interpretation.
Don't design the schema up front. Start with three plain sections, current truth, in-flight work, blockers, make reading them mandatory at the entry point, and let real friction over a few weeks reveal what the format actually needs to become.
Bottom Line: A multi-agent pipeline only remembers what someone forced it to write down and forced the next agent to read.
Enjoy this article?
Listen to the Claude Code Conversations radio show or join the community.
