Shared memory patterns and the poisoning path
Full message pool (AutoGen, MetaGPT)
agent A
agent B
agent C
pool
append-only log
every agent reads all
Blackboard with topics (CA-MCP, Matrix)
publisher
publisher
prices
alerts
summaries
subscriber
subscriber
subscriber
memory poisoning — the failure mode shared memory invents
retriever
hallucinates decimal
summarizer
echoes 42%
analyst
recommends adoption
final report
42% — never existed
verifier
read-only
re-fetches sources
flags inconsistency
output routed to
a separate channel
mitigations: append-only log · provenance fields · unwritable verifier · independent source access