# RE-call

Agent memory built on a developer's own PostgreSQL and pgvector, where every retrieved result carries a verdict, a confidence score and provenance, and a query that cannot clear a calibrated trust threshold is refused instead of answered from the nearest match.

RE-call is an agent memory system that runs on a developer's own PostgreSQL database with the pgvector extension. It indexes source documents and keeps validity, lineage, confidence and provenance attached to every stored result, rather than treating vector search alone as the answer.

Because plain vector search returns nearby text without checking whether it is current or trustworthy, RE-call marks a superseded claim as such and returns an explicit abstain response, with a reason, when no result clears its calibrated trust gate. The default retrieval path combines local embeddings with hybrid dense and PostgreSQL full-text search and needs no large-language-model call; optional modules add further retrieval, reasoning and structured-fact extraction, including a bounded one-hop graph expansion that still has to pass the same trust boundary before it can support an answer.

It is reachable from a command-line interface and from an MCP tool, so agents such as Claude can query it directly as a memory layer rather than an unverified scratchpad.

Website: https://giulioder.github.io/RE-call/