What if LLM memory isn’t just about remembering?
I have spent a lot of time working with AI across different projects, and I noticed that even when a model had access to parts of an older conversation, it did not necessarily seem to understand the significance of what had happened before. I could return to something I had worked through weeks earlier and find myself explaining the same decision again, because the information might still be somewhere in the conversation history, but the context around why that decision was made was no longer there in a useful form.
While working on my Master’s thesis, I have been spending an unreasonable amount of time reading papers about how AI systems learn, remember, and adapt. Somewhere in that reading, I came across a paper arguing that episodic memory could be an important missing piece for long-term LLM agents. The idea was that agents need to retain particular experiences, not just facts, if they are expected to operate over long periods.
That sent me down a much deeper rabbit hole.
What I found was that the research has already moved well beyond simply storing previous conversations. The question became less about whether an agent can remember and more about what happens to an experience once it becomes a memory.
The field now has several ways of giving agents persistent memory. Systems that store conversations, retrieve previous experiences, summarise interactions, update memories, remove them, and use past trajectories to influence what they do next.
Google’s ReasoningBank for example, takes successful and failed agent trajectories and distils them into reusable reasoning strategies. The agent can then retrieve those strategies when it encounters similar problems, creating a loop between experience, memory and future performance.
Anthropic has approached the problem from a more practical direction in its work on long-running Claude agents. Its agents maintain a persistent progress file containing discoveries, failed approaches, current status and known limitations so that subsequent sessions can continue the work.
Looking at these systems, I stopped thinking about memory as simply a way to keep information available.
The more interesting question became: What happens to an experience when an agent turns it into something it can reuse?
From an experience to a strategy
Human memory gives us a useful way to think about this.
Psychology distinguishes between episodic memory, which involves particular experiences, and semantic memory, which captures more general knowledge. An experience can therefore become something more abstract than the event itself.
I think about this whenever I notice how people, including me, learn from repeated interactions. A particular outcome can change an expectation about what will happen next. The memory probably is not a perfect recording of the original event. What stays with you is some combination of what happened, what you noticed about it, how you interpreted it, and what you decided to do differently afterwards. A single experience can become a rule.
Human behaviour is shaped not simply by what we experience but by the expectations and beliefs that experiences leave behind. We constantly generalise from individual events, often without being conscious that we are doing it.
Agent memory is beginning to work in a similar direction. It can retain the trajectory of an interaction, but at some point it has to decide what that interaction means for what it should do later.
An agent might retain :
“I tried approach A on this task, and it failed.”
Or it might turn several experiences into :
“For tasks like this, approach B tends to work better.”
The second is easier to reuse. It is also further removed from the evidence that produced it.
That distinction is becoming important in current research.
A 2026 study, Useful Memories Become Faulty When Continuously Updated by LLMs, found that repeatedly consolidating experiences into higher-level memories could improve performance initially and then degrade it. In some experiments, the resulting memory eventually performed worse than having no memory, while retaining the original episodic traces remained competitive. The authors argue that raw experiences should remain available as evidence rather than being continually overwritten by abstraction.
So the problem is not simply whether an agent can remember an experience. It is whether the generalisation extracted from that experience is still a good representation of what happened.
When memory becomes learning
Recent work is starting to treat memory management itself as something an agent can learn.
Get Manjusha Chandran.A’s stories in your inbox
Join Medium for free to get updates from this writer.
We learn from outcomes by updating expectations. A single outcome can influence future behaviour, while repeated outcomes can make a pattern feel increasingly reliable. But generalisation has a limit. An experience can be informative without being representative.
- If one approach fails, how much should that change what the agent does next time?
- If it succeeds repeatedly, when should that become a reusable strategy?
- If the environment changes, how should the agent recognise that an old strategy no longer applies?
Research is beginning to treat these decisions as part of the memory system itself. MemCon, for example, published in July, models memory operations as a learned decision process, adapting when and what to retrieve, when to reuse plans, and when to consolidate or forget.
Other work such as Amory, published at EACL 2026, takes another approach, organising conversational fragments into episodic narratives while semanticising peripheral facts into more general memory.
And a paper published this month, CoEvo-Mem takes this further by looking at the interaction between retrieval and the memory bank. What an agent retrieves affects which memories receive feedback, while changes to the memory bank affect what gets retrieved next time. The result is a feedback loop between memory and future behaviour.
Taken together, these approaches suggest a shift:
experience → memory → retrieval → action → outcome → memory
An experience contains context. A strategy removes some of that context so it can be applied elsewhere. That abstraction is what makes memory useful, but it also creates the possibility of generalising too far.
As agents accumulate longer histories and increasingly learn from their own trajectories, the quality of that transition may matter as much as the amount of memory they have.
I want to keep exploring this space where AI, behaviour science, and how systems learn from experience overlap.
- How an agent's experiences become knowledge that changes what it does?
- How much context should stay with a memory?
- How many experiences are enough before something becomes a pattern?
- And what happens when the situation changes?
References-
- Attentional switch to memory: An early and critical phase of the cognitive cascade allowing autobiographical memory retrieval-https://link.springer.com/article/10.3758/s13423-023-02270-w
- ReasoningBank: Enabling agents to learn from experience -https://research.google/blog/reasoningbank-enabling-agents-to-learn-from-experience/
- Rethinking Memory in AI: Taxonomy, Operations, Topics, and Future Directions-https://arxiv.org/html/2505.00675v1
- Useful Memories Become Faulty When Continuously Updated by LLMs-https://arxiv.org/abs/2605.12978
- CoEvo-Mem: Co-Evolving Retrieval Policy and Memory Bank for LLM Agents-https://arxiv.org/html/2608.01739v1
- Memory as a Controlled Process: Learned Adaptive Memory Management for LLM Agents-https://arxiv.org/abs/2607.13591
- From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms-https://aclanthology.org/2026.findings-acl.2069/
- Position: Episodic Memory is the Missing Piece for Long-Term LLM Agents-https://arxiv.org/pdf/2502.06975?
- Memory OS of AI Agent-https://arxiv.org/pdf/2506.06326
- https://cognitiontoday.com/memory-models-in-psychology-understanding-human-memory/
