AkbasCore MAM: Frozen LLM Achieves 72/72 Accuracy on Sealed Benchmark Using Source-Free Numerical Memory Cartridges

Core Mechanism: DC6 Consolidation

AkbasCore MAM introduces a third paradigm for LLM memory, distinct from fine-tuning and RAG. It enables a frozen Mistral-7B-Instruct-v0.3 to gain persistent, incremental memory by plugging in numerical "cognitive cartridges."

• Cartridge Creation: A fact is processed once through the frozen model. The system saves the hidden state at layer 6 (H6) and the attention keys/values for layers 0–6. The original text is discarded.
• Injection (DC6): During inference, layers 0–6 are skipped. The stored H6 state is injected directly into the model's internal state, and layers 7–31 run normally, allowing the new cartridge to consolidate with existing memory.
• Position Correction: RoPE rotations are mathematically adjusted to place cartridges at new positions without approximation.

Benchmark Results (TEST560)

The method was tested on a sealed panel of 24 synthetic worlds with 4 relation types, resulting in 72 test cases.

| Condition | Source Text in Input? | Accuracy | | :--- | :--- | :--- | | MAM, append-only (INCR_DC6) | No | 72/72 | | MAM, batch (BATCH_DC6) | No | 72/72 | | Model reads full text (JOINT) | Yes | 72/72 | | Naively stacked cartridges (INDEP) | No | 13/72 |

The append-only method matched the performance of the model reading the full source text, while naive stacking failed significantly (13/72).

Verification and Constraints

• Integrity: Model weights remain unchanged (verified via SHA-256 hashing). Previous memory rows are bit-exact identical after new appends.
• Scope: This is a proof-of-concept release (mam-v1.0.0), not a production-ready product. It is currently limited to 5-cartridge memories on a single model.
• Efficiency: Each cartridge is compact (~36 KiB per token) compared to a full KV cache (~128 KiB per token).
• Availability: Code and logs are available on GitHub, with a DOI provided for the release.