Claude Opus 5 really disappointed me with this result.
I had it attempt to create the same type of educational 3D video I had Fable 5 create (see the quoted post) and the result is definitely a caliber below.
Watch both, you'll see what I mean.
A senior Anthropic engineer leaked his internal PDF on giving multi-agent systems permanent memory.
Your agent forgets everything the second its context window closes. His fix: a knowledge graph that survives past any single run.
Extract → Resolve → Assemble → Query → Repeat:
Extract: Haiku pulls entities and triples, one call per doc
Resolve: Sonnet matches "Edwin Aldrin" to "Buzz Aldrin" with zero string overlap, just context
Assemble: canonical nodes, typed edges, provenance on every triple
Query: Sonnet reasons over the graph, every answer points to a specific edge
Wire it in as shared memory: workers write to it, evaluators check claims against it, loops run overnight without losing state.
Save this before it gets taken down 👇