Memory-cartridge architecture replicates 72/72 on frozen Qwen2.5-7B, no fine-tuning
Nearby_Indication474 · reddit · 2026-10-09
AkbasCore MAM's "memory cartridge" architecture, previously shown on Mistral-7B, now transfers to a completely different family: fully frozen Qwen2.5-7B-Instruct, with no fine-tuning, no RAG, and weights verified unchanged via hashing.
Mechanism
- Each fact passes through the frozen model once; the model's internal state snapshot is saved as a memory cartridge
- Cartridges stack append-only; the original text can be discarded, and the model answers from cartridges alone
- The key "cut" layer is measured per model family: after layer 6 for Mistral, after layer 3 for Qwen
Results
- Sealed test without source text: 72/72 on Qwen2.5-7B, matching the model reading the full text; naive memory concatenation scores 26/72
- Qwen cartridges are compact: 15 KiB/token vs 56 KiB/token for its full cache (Mistral: 36 KiB/token)
Author stresses this is a proof demonstration, not a product; Mistral data is in the v1.0 record (DOI 10.5281/zenodo.23245358). Open-source demo available.
Related event: Frozen 7B Models Gain Persistent Memory via "Memory Cartridge" Architecture(2 posts)→
More from coding & agent
- Neon ships official AI Gateway adapter for TanStack AI with single-credential multi-model access — matei_zaharia · 2026-10-23
- Modal open-sources Spindle, a Tinker-compatible training server you pay for in GPU time — andersonbcdefg · 2026-10-23
- Reflection AI's Beam agent auto-builds a Gemma-4 fine-tuning notebook, boosting accuracy 66.5% — danielhanchen · 2026-10-23
- Cua Spaces SDK lets Hermes and other agents run in sandboxes on your own hardware — Teknium · 2026-10-23
- PHP hit by two unauthenticated RCE flaws; ProjectDiscovery launches free internet-wide alert service — cyb3rops · 2026-10-23
- Modal open-sources Spindle, a Tinker-compatible training server that charges GPU time instead of tokens — josh_wills · 2026-10-23