Hello. I'm sharing a side project I created while unemployed.
Some of you may have seen the movie "Memento," where the protagonist can't retain short-term memories and lives by tattooing facts on his body. I wondered what would happen if I applied that concept to a small LLM (Gemma 4 E4B). The name is Gemento.
The core question was simple - instead of scaling up the model, what if I kept the same model but called it multiple times with structured external memos (tattoos)? Could we achieve complex reasoning without a larger model?
The most interesting result (multi-step loop quality accumulation):
Same Gemma 4 E4B, same problem, only number of calls varies:
* Single inference : 50.0%
* 2 loops : 72.2%
* 4 loops : 88.9%
* 8 loops : 94.4%
The model didn't change, only the orchestrator (loops + external memos) was refined, yet the accuracy monotonically increased. I was frustrated when v1 initially returned 0%, but design is key.
Other validated findings:
- Attaching a calculator, linear algebra, LP solver transformed hard math problems from 0% → 100%
- For 3-hop questions on long documents, role separation (ABC) + chunk iteration outperformed RAG/full injection (interestingly, compared to large LLMs, "lost in the middle" seems to decrease dramatically - though with computation time trade-off)
The environment is Windows + LM Studio + Gemma 4 E4B, so a regular PC is sufficient. MIT license, along with 9 experiment notes are included.
GitHub: https://github.com/hang-in/gemento
Research Notes
https://github.com/hang-in/gemento/blob/main/docs%2Freference%2FresearchNotebook.md
I'd appreciate if you share similar research directions. Feedback and issues are welcome.
** I'm infinitely grateful to the local LLM users of Damo-Ang AI Party who inspired this experiment. **
