Shared Memory Architecture

Definition

A multi-agent pattern that decouples state from invocation: subagents index their outputs into a shared vector store, and subsequent agents use semantic search to retrieve only the relevant prior findings instead of passing full logs around.

Key points

• Anti-pattern: Daisy-chaining full conversation logs between subagents — this scales token costs exponentially.
• Architect's pattern: Decouple state from invocation. Subagents index their outputs into a shared vector store; later agents retrieve only relevant prior findings via semantic search.
• Prevents state loss when a multi-agent pipeline crashes mid-processing.
• Diagram: Web Search Agent (writing findings) → Shared Vector Store → Document Analysis Agent (reading findings via semantic search).

Why it matters for the exam

• This is the Multi-Agent × Token Bloat cell in the Reference Matrix ("Shared Vector Store").
• Tests whether you know to persist state externally rather than daisy-chaining logs, and that semantic search retrieves only relevant slices.

Common gotchas

• Passing full logs between agents is the wrong answer — it is the named anti-pattern (exponential token cost).
• The benefit is not just cost: it also provides crash resilience (state survives a mid-pipeline crash).

See also

• Hub-and-spoke orchestration
• Scratchpad Pattern
• Structured Intermediate Representations
• Architect's Reference Matrix

Sources

• architects-playbook