soma-memory-fix

To fix unwired kb_search nodes, you audit production AgentStack agents to locate nodes missing knowledgeBaseId, match them to correct knowledge bases in a strict 1:1 manner, and apply patches using live MCP tool responses. The process inspects flows, proposes a plan, and verifies post-patch outcomes.

Safely patching agent memory lookups requires reading current node states via as_inspect_flow, matching unwired kb_search nodes to knowledge bases, applying changes with as_patch_node_field using only live IDs, and re-reading the flow to verify the field is set before marking it fixed.

You need to audit agents for missing knowledgeBaseId when production AgentStack deployments contain unwired kb_search nodes that fail to retrieve memory. This occurs across multiple agents and requires mapping each node to its correct knowledge base safely.

Yes, the kb_search patching process works across multiple production agents simultaneously. It audits each agent flow individually, lists available knowledge bases, and ensures a strict 1:1 KB mapping for each unwired node before applying any patches.

Agent memory patching halts when knowledge base matching is ambiguous, meaning there are zero or two plus KBs available for an agent. This strict constraint prevents guessing and ensures every patch uses a correctly matched, live knowledge base ID.

To verify a successful knowledgeBaseId patch without hallucinating IDs, you must perform a post-patch re-read using as_inspect_flow to confirm the field is set. Every patched ID must originate from a live MCP tool response in the current session, never memory.