Buzo

Observability and integrity for the vector memory of AI agents in production. Understand what your agents know, what they are retrieving, and whether that knowledge can be trusted. Attribute when flagged content reaches a real user. Every action is human-gated, auditable, and reversible.

Buzo connects to your vector store (Pinecone, ChromaDB, pgvector, Qdrant), runs a deterministic analysis pipeline combined with AI inspectors over your corpus, and surfaces findings in a dashboard where every write against your store is gated by an explicit human action with a typed confirmation keyword.

Install buzo-sdk once, and Buzo also starts seeing what your agent actually consumes — turning static findings into usage-weighted priorities,SERVED_FLAGGED alerts, and CITED_FLAGGED attribution when a flagged vector ends up in the agent's final response.

Start here

What you get out of the box

• Usage-weighted findings. Instead of a flat list of issues, see which problem vectors are actually being served to users and how often.
• SERVED_FLAGGED alerts. Email within minutes when your agent retrieves a vector Buzo previously flagged for review.
• CITED_FLAGGED attribution (v0.2+). The strictly stronger signal — the flagged vector's content actually appears in the user-facing answer, not just in the context.
• Tamper-evident audit chain. Every action against your store — quarantine review, sync-apply, MCP pass-through — is logged with a SHA-256 hash chain.