MATLAB Memory Optimization Workflow

Systematic 7-step workflow for finding and fixing memory bottlenecks in MATLAB code.

When to Use

• User gets out-of-memory errors running MATLAB code
• User wants to reduce memory usage of their MATLAB program
• User wants to process larger datasets without running out of memory
• User asks to profile or measure memory allocations

When NOT to Use

• The bottleneck is execution speed, not memory (use matlab-optimize-performance)
• The memory issue is in compiled C/MEX code that can't be changed at the M-code level
• Memory usage is dominated by I/O buffers (memory-mapped files, database connections)

The 7-Step Workflow

Step 1: Establish Memory Baseline

Measure current memory usage before making changes.

m0 = memory;
targetFunction(inputs);
m1 = memory;
deltaBytes = m1.MemUsedMATLAB - m0.MemUsedMATLAB;
fprintf('Memory delta: %.2f MB\n', deltaBytes / 1e6);

When memory errors (Linux/macOS), use whos for variable sizes or Java runtime for heap:

info = whos('result');
fprintf('Variable size: %.2f MB\n', info.bytes / 1e6);

Step 2: Profile Memory Allocations

Find where memory is being allocated.

profile('-memory', 'on');
for iter = 1:5
    targetFunction(inputs);
end
profile off;
p = profile('info');
ft = p.FunctionTable;
[~, idx] = sort([ft.TotalMemAllocated], 'descend');
for i = 1:min(15, numel(idx))
    f = ft(idx(i));
    fprintf('%-40s %10.2f MB\n', f.FunctionName, f.TotalMemAllocated/1e6);
end

If TotalMemAllocated fields are zero, fall back to whos snapshots before/after each function call.

Key things to look for:

• Functions with high "Allocated" but low "Freed" — memory is retained
• Functions called many times with moderate allocations — total adds up
• Large gaps between Allocated and Freed — temporaries accumulating

Step 3: Identify Optimization Opportunities

Based on profiling, identify which patterns apply. See references/memory-patterns.md for code examples.

Pattern	Typical Reduction	Look For
Cell collection + vertcat	O(N²) → O(N)	[arr; newRow] inside loops
Implicit expansion over repmat	Eliminates full copy	repmat(A, [1 1 K]) for broadcasting
Clear variables when done	Immediate reclamation	Large arrays used only in early steps
Break chained expressions	1 fewer peak temporary	a.*b.*c./d all alive at once
Reuse variables (overwrite in-place)	Avoids output allocation	Separate variables for each step
max/min instead of masking	Eliminates logical temporary	x .* (x > 0) pattern
zeros(...,'like',x)	Eliminates temporaries	0 * x to create zeros
Copy-on-write sharing	Shares backing memory	Same array assigned to multiple places
Dense → sparse	O(N²) → O(N·bw)	zeros(N,N) where N > 10000

Step 4: Implement Optimizations

Apply the identified patterns. Focus only on the hotspots identified in Step 2 — do not apply patterns everywhere.

Step 5: Measure Optimized Memory

Re-measure using the same method as Step 1:

m0 = memory;
optimizedFunction(inputs);
m1 = memory;
deltaOpt = m1.MemUsedMATLAB - m0.MemUsedMATLAB;
reduction = 1 - deltaOpt / deltaBytes;
fprintf('Optimized: %.2f MB (%.0f%% reduction)\n', deltaOpt/1e6, reduction*100);

Step 6: Verify Correctness

Every optimization must produce the same results:

original = originalFunction(inputs);
optimized = optimizedFunction(inputs);
maxErr = max(abs(original(:) - optimized(:)));
fprintf('Max error: %.2e\n', maxErr);
assert(maxErr < 1e-10, 'Results differ!');

Step 7: Report Results

Summarize the memory optimization with baseline, optimized, reduction percentage, correctness check, and patterns applied.

Key Rules

1. Never propose optimizations based solely on reading source code — always measure and profile first
2. Verify correctness — memory optimizations must produce identical results
3. Clear variables early — free memory as soon as data is no longer needed
4. Avoid growing arrays — preallocate or use cell collection
5. Break chains — sequential assignment reduces peak memory vs chained expressions
6. Watch for copies — MATLAB copies on write; reuse variables to avoid duplicates

Platform Notes

• Windows: memory command returns full statistics (MemUsedMATLAB, etc.)
• Linux/macOS: memory errors ("not supported on this platform"). Use whos for variable sizes, Java Runtime.getRuntime for heap usage, or OS-level RSS via system('ps ...')
• profile -memory: Works on all platforms but is undocumented since R2016a. When unavailable, use whos snapshots before/after function calls.

Copyright 2026 The MathWorks, Inc.