RFC: Memory

[feynon]

Experiment 1

Status: Active

Solving memory with a fully fine tuned model trained for memory operations.

Implementation

We plan to fine-tune two models as our first in-house memory models: a larger model for standalone agent use cases, and a smaller model where the agent operates purely as a memory module.

Dria’s research, code, and models are open source, making them a strong starting point. Below is a rough implementation plan for Tiles:

1. Reproduce Dria’s mem-agent to develop a solid understanding of the memory framework.
2. Use Unsloth to train a new base memory model using Unsloth base weights.
3. Fine-tune a LoRA-based memory extension with Unsloth on a custom dataset.
4. Learn the developer experience of Tinker for LoRA fine-tuning.
5. Port the Tinker workflow into Tiles using the Modelfile SDK.

---

Experiment 2

Status: Proposed

Implement Recursive Language Models using conversational chat history grounded in code execution, with progressive disclosure via text compaction and retrieval at test time using base oss models.

Implementation

A combination of code sandbox, skills integration, and DSRs based prompt optimizer. Being worked on #55, https://github.com/tilesprivacy/tiles/issues/56, #45, and #66.

---

References

• https://dria.co/research/mem-agent:-equipping-llm-agents-with-memory-using-rl
• https://github.com/firstbatchxyz/mem-agent
• https://www.mcpjam.com/blog/claude-agent-skills
• https://arxiv.org/html/2512.24601v1