Use memory graphs to prove why memory survives a defined lifetime boundary. Separate unreachable leaks from reachable growth, preserve raw tool output, and verify the same app-owned type and ownership path after a fix.

This skill owns capture and command-line ownership/growth analysis. Use the Memory Graph Debugger or Instruments when their interactive graph and allocation timeline are the primary task. Use source review for a suspected closure capture only after runtime evidence identifies the lifetime or path.

Do not collapse these conditions:

Apple's leak scanner uses conservative pointer discovery and incomplete type metadata. Counts can fluctuate, and a zero result does not prove the absence of an ownership bug. Strong evidence identifies the expected lifetime, an app-owned type or allocation, and a credible path or isolated reproduction.

Name the object that should disappear and the event that ends its useful life. For example: should deinitialize after dismissing the editor and completing pending save work.

Record one deterministic sequence:

Keep build, simulator/device, data, Malloc Stack Logging setting, and repetition count stable. Malloc Stack Logging adds valuable allocation backtraces but also overhead; compare only runs with the same setting.

Xcode can export a graph from the Memory Graph Debugger. For a running Simulator app, use the helper from this skill:

mkdir -p /tmp/myapp-memory
mkdir /tmp/myapp-memory/run-01
python3 scripts/capture_sim_memgraph.py \
  --bundle-id com.example.MyApp \
  --output-dir /tmp/myapp-memory/run-01 \
  --pretty > /tmp/myapp-memory/run-01/capture.json

The per-run must fail if the capture directory already exists. Use a new run name rather than mixing stale evidence with a retry.

Pass when more than one Simulator is booted. The helper accepts only one exact launchd label and PID; zero or multiple matches are errors. It runs the host command, retains stdout/stderr, and writes a manifest. Do not replace this with or a substring match.

Capturing suspends the process. Do not use capture latency as performance data.

MEMGRAPH=$(jq -er \
  'select(.status == "captured") | .memgraph | select(type == "string" and length > 0)' \
  /tmp/myapp-memory/run-01/capture.json)
test -s "$MEMGRAPH"
python3 scripts/summarize_memgraph.py \
  "$MEMGRAPH" \
  --artifact-dir /tmp/myapp-memory/run-01/analysis-raw \
  --app-image 'MyApp|MyFeatureKit' \
  --trace-limit 3 --group-by-type --pretty \
  > /tmp/myapp-memory/run-01/analysis.json

Read the exact graph path from the preserved capture report; do not guess a timestamped filename. The helper creates a dedicated raw-artifact directory, refuses to reuse it, runs , and parses only a conservative subset of its text. marks candidate rows; it does not prove ownership. runs bounded queries. Add when aggregate root paths are more useful than individual leaked addresses. With , that reference-tree query is grouped in the same invocation. Exit statuses 0 and 1 from remain analyzable; a primary status above 1 fails the summary, while optional-query failures are preserved and warned as unusable without discarding a valid primary summary.

Apple does not publish these text formats as stable machine schemas. Treat parse warnings as a reason to inspect the raw artifacts, not to loosen the parser until it emits a desired answer.

Start with an app-owned leaked type or allocation stack. Inspect:

An unreachable self-cycle may have no live root in . Use the grouped leak graph plus source verification or reduce the behavior to an isolated reproduction. Never invent a root path that the graph does not contain.

Use matching baseline and post-flow graphs, locate the growing region, compare object types, then trace a suspicious address back to an app-owned edge. The evidence goal is persistent reachable growth across the same lifetime—not a lower RSS value or a single large snapshot. Load reachable-growth.md only for this empty-leak branch; it contains the ordered , , , and queries and their logging-dependent alternatives.

Prefer the narrowest ownership correction: break the unintended strong edge, cancel work that owns the object, remove an observer, bound/evict a cache, or release a large buffer after its last use. Use when the reference may legitimately become ; use only with a proven lifetime guarantee.

Repeat the identical flow. A fix is supported when the same app-owned type/path disappears or the pre/post growth attributable to it is removed across repeated runs. Lower RSS, a smaller graph file, or a lower aggregate leak count alone is not proof.