memoception
Game studios

Unreal Engine optimization

A headless UE5 optimization loop with a persistent brain.

−38%
draw calls after 3 passes
−22%
avg frame time
time series
profiling, queryable
1 brain
across every pass
How it fits together

   ┌─────────┐    store · recall     ┌──────────────┐
   │ Claude  │ ◄───────────────────► │  Memoception │
   └────┬────┘                        │   (memory)   │
        │  MCP                        └──────┬───────┘
        ▼                               profiling, logs,
   ┌──────────────┐    cook · trace          conventions
   │  Unreal MCP  │ ◄── headless harness ──────────┘
   │  (your VPS)  │      optimization passes
   └──────┬───────┘
          ▼
   ┌──────────────┐
   │  UE5 project │  ── assets, config, source ──► stored & contextual
   └──────────────┘

The challenge

Large UE5 projects lose all context between optimization passes — profiling traces, prior fixes, engine conventions, and thousands of assets across a dozen formats live nowhere the agent can reach. Every pass starts cold.

How Memoception solves it

1
Memory becomes the project's brain

The .uproject, Config/*.ini, C++ source, Blueprints, and profiling CSVs are custom-stored with rules — profiling as a queryable time series, everything tagged and recallable.

2
Claude drives Unreal on your VPS

Claude connects to an Unreal MCP running headless on your VPS, so it can open the project, cook, and capture Unreal Insights traces without a workstation in the loop.

3
The headless harness runs the passes

Our Unreal headless harness executes cook + Insights + automation jobs, captures the logs, and stores them — so every run's evidence accumulates instead of vanishing.

4
Recall conventions per bottleneck

For each detected bottleneck (draw calls, Lumen/Nanite GPU, texture pool, PSO hitches) Claude recalls the relevant UE conventions and this project's own prior findings, then proposes and generates fixes (scalability presets, LODs).

5
Every asset, every format

All game assets are stored and made contextual — meshes' metadata, materials, data tables, configs — get context from any format via the custom-store pipelines.

Built with
custom/store (ini, cpp, timescaledb)memory/recallUnreal MCP (headless)headless harnessorganize → SQL