Research.
Coding agents drift toward generic answers because their context resets between runs. Sigilix research is about the fix: memory-native agents backed by an org and developer memory index, grounded review, and repair loops that keep hold of local truth.
TL;DRNotes on memory-native coding agents: why context is not memory, and how grounded review becomes repair.
Research on memory-native agents and grounded review
Research note
When models fight memory
Why context-memory conflict shows up in coding agents, why generic retrieval is not enough, and how Sigilix turns memory into product infrastructure.
Research note
Context is not memory
Why larger windows are the wrong abstraction for coding agents, and what durable repo memory should do instead.
Research note
From review to repair
A product note on why review findings need enough proof, memory, and verification to become safe repair work.
Frequently asked questions
- What does Sigilix research focus on?
- Memory-native coding agents: why a bigger context window is not the same as memory, what happens when a model's broad prior fights a repository's local truth, and how a grounded review can become safe repair.
- Is context-memory conflict a Sigilix term?
- No. Knowledge conflict, and specifically context-memory conflict, come from the research literature: they describe external context disagreeing with what a model learned in its parameters. Our notes explain why this shows up in coding agents and how we design around it.
- Do these notes report benchmark claims?
- Where numbers appear they are scoped to the reference fixtures we disclose, and any lift is described as an org- or user-scoped memory effect, not a product-wide training claim. The notes are about architecture and behavior, not leaderboard ranking.