Can an AI get smarter just by how it writes its own notes, with no retraining at all?
How does textual memory structure affect frozen model improvement?
This explores whether the way an AI agent writes down and organizes its memories in plain text, rather than changing the model's weights, can make a frozen model keep getting better.
This explores whether the shape of an agent's written memory, and not just how much it remembers, decides whether a frozen model keeps improving. The corpus says shape matters a lot. Agents that store memory in a causal form, meaning each lesson keeps the conditions under which it applies, beat generic reflection by 23 points on repeated trials. They also gain 4-17 points when moved to new environments (Can frozen language models continually improve through memory structure alone?). A note like "this worked" is close to useless. A note like "this worked because X was true" tells the model when to reuse it.
A second line of work reaches the same conclusion from another direction. Pulling natural-language rules out of a body of context and saving them as reusable skills lifted GPT-4.1 on CL-bench from 11.1% to 16.5%. Those skills also transferred across different model backbones (Can frozen models learn better by extracting context into skills?). The gain comes from the model's own knowledge being distilled into a more usable form, not from adding new information. The memory is written for reuse rather than as a transcript of what happened.
Structure also works at a larger scale. One long-horizon harness arranges state in four levels: the weights, the context window, a persistent workspace with subagents, and disk-backed history. That lets a model read and write addressable state beyond a single prompt (Can external state caches let models solve harder problems?). The authors report gains on ARC-AGI-3 and say the setup separates harness failures from model failures. That separation is useful, because a bad result can be blamed on the memory layout instead of the model. The caveat is that they didn't test which components actually drive the gains.
The cautionary result is that structure can also hurt. On MemTrapBench, all five memory frameworks tested did worse than having no memory at all, by more than 10%, even though the memories were accurately stored and relevant (Can relevant memories actually harm LLM reasoning?). Accurate recall isn't enough. The failure shows up when the model uses the memory, which is why a format that keeps applicability conditions, as in the causal-memory work, matters.
The alternatives put the memory somewhere other than text the model reads. One trains a separate memory model that works with frozen proprietary models and avoids search costs that grow with corpus size. The price is up-front training and limited capacity (Can a separate memory model inject knowledge without touching the LLM?). Another builds memory into the model backbone, arguing that external memory and the model can optimize against each other (Should agent memory live inside the model backbone?). Both give up the appeal of textual memory, which is that you can read it, edit it, and carry it to a different model. The corpus doesn't settle which approach wins, but it does show that with a frozen model the memory's shape is one of the few levers you have.
Sources 6 notes
Agents using causal-form memory (preserving applicability conditions) outperform generic reflection by 23 points on repeated trials and gain 4-17 points transferring to new environments, showing memory shape matters more than parameter updates.
Extracting natural-language rules from context into reusable skills improves frozen model reasoning without weight updates. On CL-bench, this lifts GPT-4.1 from 11.1% to 16.5%, with skills transferable across model backbones.
Prime Agent organizes persistent state in four levels (weights, context, persistent REPL plus subagents, disk-backed history) to let models read and write addressable state beyond their instruction stream. The approach isolates harness failures from model failures and reported gains on ARC-AGI-3, though specific components remain unablated.
MemTrapBench shows that all five tested memory frameworks underperform a no-memory baseline, with drops exceeding 10%, despite memories being accurately stored and task-relevant. This reveals a failure mode at the point of use that standard memory benchmarks miss.
MeMo trains a dedicated memory model to encode new knowledge, eliminating inference-time search costs that scale with corpus size. It avoids fine-tuning risks and works with frozen proprietary models, but trades this for up-front training cost and capacity limits.
Show all 6 sources
Metis demonstrates that agent memory can be implemented as a persistent state and autonomous procedures within the model backbone rather than external modules. This approach enables end-to-end training and avoids the decoupling failures where external memory and backbone optimize independently.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Know It, Act on It: Investigating Memory Utilization in LLM Personalization
- Useful Memories Become Faulty When Continuously Updated by LLMs
- GateMem: Benchmarking Memory Governance in Multi-Principal Shared-Memory Agents
- MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use
- Metis: Memory Foundation Model
- Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory
- MeMo: Memory as a Model
- Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering