Skip to content
Critique/docs
Architecture

Memory & Context System

How Critique remembers repository context for chat, reviews, and optional promoted learnings.

Critique is repository-aware. Automated reviews always use fresh PR evidence; memory helps chat and follow-up sessions recall decisions and facts across time.

Layers of context

LayerWhat it provides
Session contextWhich repo, PR, or review run you are working on
Repository indexSearchable code chunks and embeddings for grounded answers
Long-term memoryCurated board entries and chunks promoted when useful

Repository indexing

When you connect repositories, Critique can index code for semantic and keyword search. Indexing refreshes after installs and significant default-branch updates. Indexing consumes credits in the background; see Usage.

Chat tools such as repository search prefer warm index results and may fall back to GitHub code search when needed.

Hybrid retrieval

When you ask about prior decisions or “what did we learn last time?”, Critique blends:

  • Keyword overlap on stored notes and chunks
  • Semantic similarity on embeddings

Scope is limited to namespaces you have access to (repository, review run, or chat session). This avoids treating unrelated repos as one memory pool.

Promotion (curated memory)

Not every review is copied into long-term memory automatically. Promotion happens when a human (or an explicit agent step after confirmation) decides a run produced durable guidance worth keeping for future PRs.

If you expected automatic learning from every merge, that is intentional: unfiltered promotion would pollute memory with noisy or wrong runs.

Remedy successes can also be promoted when operators want fixes to inform later work.

Privacy

Only repositories under your GitHub App installation are indexed. Chat history retention can be limited via account settings where offered.