Storing an AI's memory outside its brain sidesteps forgetting and corruption, but does it just move the problem to recall time?
Does moving memory outside model weights avoid the limitations of in-weight retention?
This explores whether keeping knowledge and experience outside a model's weights (in a separate memory model, an episodic store, or a cache) gets around the problems of storing it in the weights, such as forgetting and corrupted knowledge.
This explores whether keeping knowledge and experience outside a model's weights gets around the problems of storing it in the weights, such as forgetting and corrupted knowledge. The corpus says it avoids the damage that weight updates can do, but it moves the difficulty to a new place: the moment the model has to use what was stored.
What external memory avoids is real. Direct fine-tuning can corrupt knowledge that lives in a model's lower layers, while Can decoding-time tuning preserve knowledge better than weight fine-tuning? leaves the base weights alone, closes most of the alignment gap, and beats fine-tuning on knowledge tasks. Can editing hidden representations beat weight updates for finetuning? makes a similar move by learning small interventions on frozen representations. Neither is a memory system, but both show the appeal of not touching the weights. Memory systems apply the same idea. Reflexion agents write a plain-language diagnosis after each failure and improve across episodes with no parameter updates Can agents learn from failure without updating their weights?. AgentFly reaches 87.88% on GAIA validation purely through reading and writing case, subtask, and tool memories Can agents learn continuously from experience without updating weights?. A dedicated memory model can even feed knowledge to a frozen proprietary LLM without search costs that grow with the size of the corpus Can a separate memory model inject knowledge without touching the LLM?.
What it doesn't avoid is failure at the point of use. In weights, the risk is that storage gets damaged. Outside weights, the risk is that accurate storage still doesn't help the model. MemTrapBench found that all five memory frameworks it tested did worse than having no memory at all, by more than 10%, even though the memories were faithfully recorded and relevant to the task Can relevant memories actually harm LLM reasoning?. External memory also has costs of its own. The separate memory model needs up-front training and has capacity limits, and Reflexion works only because success/failure feedback is unambiguous and the reflections are kept uncompressed, which means the store keeps growing. The question of what to keep and what to forget doesn't go away either. One result shows most intermediate reasoning tokens quickly lose importance, so a short prefix plus a recent window is enough for a 3x speedup Can models think longer by forgetting intermediate reasoning?.
The line between inside and outside is also blurrier than the question assumes. Prime Agent treats the weights as just the first level of a four-level state hierarchy, followed by context, a persistent REPL with subagents, and disk-backed history. That layout lets its authors separate harness failures from model failures Can external state caches let models solve harder problems?. Titans puts a learned neural memory next to attention that stores surprising tokens and scales past two million tokens of context, so the memory is trained rather than a plain lookup store Can neural memory modules scale language models beyond attention limits?.
Some work argues external memory is a holding area rather than a replacement. The Sleep paradigm says continual learning still needs to consolidate in-context knowledge into weights, using upward distillation and RL-generated rehearsal, and that this can be done without forgetting Can models consolidate memories during offline sleep phases?. Put together, the corpus suggests that moving memory out of the weights protects what the model already knows and lets an agent learn cheaply from experience. It doesn't remove the need to decide what is worth remembering, to retrieve it well, or, over the long run, to consolidate the most durable knowledge back into the weights.
Sources 10 notes
Proxy-tuning closes 88-91% of the alignment gap while surpassing direct fine-tuning on knowledge tasks by leaving base model weights untouched. Direct fine-tuning corrupts knowledge storage in lower layers, whereas proxy-tuning applies distributional shifts that primarily affect reasoning and style.
ReFT learns task-specific interventions on frozen model representations rather than updating weights, with LoReFT (low-rank linear subspace variant) dramatically outperforming LoRA across reasoning, instruction-following, and NLU benchmarks while using far fewer parameters.
Reflexion demonstrates that unambiguous environmental feedback (success/failure) enables agents to write useful self-diagnoses and improve across episodes without parameter updates. The binary signal prevents rationalization, and keeping reflections uncompressed preserves their usability.
AgentFly formalizes agent learning as a Memory-augmented MDP with three memory modules (case, subtask, tool) that enable credit assignment and policy improvement entirely through memory operations. The approach achieved 87.88% on GAIA validation without modifying LLM parameters.
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 10 sources
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.
Most intermediate reasoning tokens become unimportant as reasoning progresses, so keeping only the instruction prefix and a recent window achieves 3x speedup without training while enabling traces beyond 100k tokens.
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.
Titans architecture separates attention (short-term, quadratic) from neural memory (long-term, compressed), prioritizing surprising tokens for storage. The model outperforms standard Transformers and linear RNNs across tasks while scaling to 2M+ token contexts without quadratic penalties.
The Sleep paradigm uses Knowledge Seeding (distilling smaller networks into larger ones) and Dreaming (RL-generated rehearsal) to consolidate in-context knowledge into weights without forgetting. Gains appear in long-context understanding, few-shot reasoning, and continual learning.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Useful Memories Become Faulty When Continuously Updated by LLMs
- Know It, Act on It: Investigating Memory Utilization in LLM Personalization
- AgentFly: Fine-tuning LLM Agents without Fine-tuning LLMs
- Memorization and Knowledge Injection in Gated LLMs
- MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use
- Language Models Need Sleep
- GateMem: Benchmarking Memory Governance in Multi-Principal Shared-Memory Agents
- Titans: Learning to Memorize at Test Time