Updated 23 June 2026

What is a canonical context layer?

Most of what people call agent memory is per-agent recall: an agent remembering its own past runs, facts, and procedures. That's useful, but it's private and it drifts — two agents can each remember a different "how deploys work." The canonical context layer sits one level up: it is the team's current truth, shared, not any one agent's recollection.

The three properties that make it canonical

• One current doc per topic. Not the top-k similar files — the one that is authoritative right now, with the stale and duplicate copies marked as such by a freshness marker.
• Shared across the fleet. Every agent and teammate reads through the same point, so there's no per-agent copy to fall out of sync. That's what makes it a single source of truth for agents.
• Written back, not just read. The layer stays current because agents write-back what they learn into it.

Write-back memory: the part that keeps it current

Write-back memory is when an agent saves what it learned back into the shared store once, through a single point, so the next agent reads it instead of re-deriving it. Without write-back, a context layer is read-only and goes stale the day it ships; with it, every solved problem becomes a record the rest of the fleet inherits. Write-back is what turns a static doc dump into a living source of truth, and it's the direct fix for re-derivation.

How trovex implements it

trovex is a canonical context layer you run locally. It indexes your docs, returns the one current section that answers a query as a path:line pointer marked canonical / stale / duplicate, and exposes a write path so agents save canonical records back through a single shared point. Serving the one current section instead of a pile of candidates is about 60% fewer tokens per lookup. SQLite + ONNX, no cloud, no API keys — your docs never leave your machine.