KARMA-LM C prototype

First portable C17 iteration of KARMA-LM: Kernel-Augmented Recurrent Memory Architecture with Loss-Managed Memory.

This prototype is intentionally small. It demonstrates:

• byte-level input
• tiny recurrent backbone
• semantic memory slots
• exact episodic memory
• rule-based compression-risk classifier
• audit/checksum hashes
• recall-before-answer behavior

The neural weights are random in this iteration. Training comes next.

Build

cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Debug
cmake --build build
./build/karma_demo

On Windows MSYS2, the executable may be:

./build/karma_demo.exe

Test

./build/test_memory

or on Windows:

./build/test_memory.exe

Next milestone

Add a tiny training loop for next-byte prediction:

• cross-entropy loss
• backpropagation through recurrent layer
• Adam optimizer
• checkpoint save/load
• synthetic memory preservation dataset

---

Version: v0.0.2-pre

What Changed in v0.0.2-pre

v0.0.2-pre improves the memory reliability layer before adding training.

New or improved behavior:

• explicit risk scoring from risk flags
• stronger exact-detail recall scoring
• exact-memory preference for risky facts
• contradiction-candidate detection
• regression tests for:
  • numbers
  • units
  • decimals
  • negations
  • code-like spans
  • audit-hash preservation
  • exact recall before answer