Engineering
Deep technical writing from the team building Recall. Not blog posts — these are the documents we wish existed when we started.
Running Recall in production: metrics, debugging, migrations, and incident response.
What to Watch When Recall Goes to Production
The three metric families that separate a healthy memory store from one degrading in silence. Prometheus scrape setup, quality metrics nobody else measures, cost attribution per pipeline stage, and three debugging workflows that give you answers fast.
20 minoperationsmetricsprometheusTracing a Memory Back to Its Source
Provenance is a first-class field on every memory in Recall, not an afterthought in logs. This page walks through how to trace a wrong memory to the conversation turn that produced it, roll back an entire bad extraction run, and debug why a specific memory didn't surface in retrieval.
16 mindebuggingprovenanceauditUpgrading Your Extractor Without Breaking Your Memory Store
Every memory in Recall carries a schema_version field. When you upgrade the extraction model or change the extraction schema, old and new memories coexist with different versions. This page covers the three migration strategies, how to detect version skew, and what a safe rollback looks like.
14 minmigrationsschemaoperations
How Recall works under the hood: pipeline stages, scoring, plugins, and data structures.