Skip to content

Using Memory in Reviews

This guide walks through the practical day-to-day usage of the memory layer.

Setup (one time)

No daemon setup required. The memory layer is zero-configuration: consolidation fires automatically in the background when you run any claude-review command and the trigger conditions are met.

Just start using --memory on your reviews:

claude-review diff --memory
claude-review pr 123 --memory

The first time you run with --memory, the .claude-review/memory.db database is created in your repo root. From that point on, the on-wake trigger fires consolidation on your next command invocation once 30 minutes have elapsed or 10+ findings have been stored.

Optional: verify DB state

claude-review memory status
Daemon: not running (not needed — on-wake consolidation is active)
Findings stored: 12 (accepted: 12)
Consolidations:  2
False positives: 0
Last consolidation: 2026-03-11 15:20

Day-to-day usage

Enable memory for every review

Add --memory to your review commands:

claude-review diff --memory
claude-review pr 123 --memory

Or set it in your config

Add to ~/.claude-review.json to enable it globally:

{
  "memory": true
}

Building up the knowledge base

Memory becomes more valuable as you accumulate more reviews. The first few reviews with --memory won't produce much context (the DB is empty), but after 5–10 reviews the memory starts making a meaningful difference.

How long until memory is useful?

Reviews completed Memory effect
1–5 Minimal — just establishing baseline
5–20 Memory starts surfacing recurring patterns in the same files
20+ Consolidation kicks in; cross-PR patterns become visible
50+ Full value — reliable cross-PR pattern detection, hotspot file awareness

Checking what's in memory

claude-review memory status

For a narrative summary of patterns:

claude-review insights

Interpreting memory-augmented results

When memory is active, the markdown report includes a note at the top:

# Code Review Report
*Memory-augmented review — context from 23 past reviews*

Individual findings from the memory context may appear as separate notes:

### ⚠️ Memory: Recurring pattern
**File**: `src/auth/token.go` has had 3 auth-related findings in the last 30 days.
This finding may be part of a broader pattern — see `claude-review insights` for details.

Team workflows

Shared team memory

Each developer has their own local ~/.claude-review/memory.db. There is currently no team-shared memory server — memory is individual.

Workaround for team sharing: Run claude-review insights periodically and share the output in your team channel or wiki. This gives everyone visibility into patterns without needing a shared server.

Pre-commit hook + memory

If you've installed the pre-commit hook and have memory enabled:

claude-review install-hook

The hook automatically uses --memory if the daemon is running, providing richer context at commit time.

Resetting memory

To clear all stored data and start fresh:

claude-review memory clear

Good reasons to do this:

  • After a major architectural refactor (old findings are no longer relevant)
  • When switching to a new project (if using a shared ~/.claude-review.json)
  • If the DB has grown large and you want to reduce noise

Troubleshooting

Findings not being stored

Make sure you're running with --memory or have it set in config. Without the flag, findings are not stored and the DB is not created.

Consolidation not firing

Run any claude-review command. If the DB exists and either trigger is met (30 min elapsed or 10+ new findings), you'll see [memory] consolidating patterns in background... on stderr.

Check the last consolidation time:

claude-review memory status

Memory not providing useful context

This is normal early on. Accumulate more reviews — context improves with volume. Consolidation only fires when there are new findings, so the first few reviews will have minimal context.