Agent Memory
Giving an AI a way to remember across conversations, since the model itself forgets everything the moment a session ends.
Reviewed July 10, 2026Contested
When not to use it
- For one-shot tasks. A translation or a summary doesn't need to remember you, and building memory into it adds privacy surface for nothing.
- When the conversation fits in the window. Re-sending it is simpler, exact, and free of retrieval bugs.
- When you can't answer "how does a user delete this?" Memory you can't erase is a liability with a UI.
Reach for something else instead
- Just re-send the conversation — context windows are large, and this is exact where retrieval is approximate.
- Explicit user profiles — a structured record the user can see and edit beats inferred memories they can't.
- Summarisation only when you need continuity within a long session but nothing across sessions.
Read more on the blog
- Agent memory: what should persist, and what should notA plain filesystem storing markdown files scored 74% on standard memory benchmarks, beating dedicated vector systems. The hard part was never storage. It is deciding what to keep and when to stop believing it.
- Why AI agents fail: the seven failure modesGartner predicts over 40% of agentic AI projects will be canceled by 2027. The failures follow patterns, seven of them. The taxonomy: what breaks, why, which real incident proved it, and which control would have prevented it.
- Context engineering: the skill that replaced prompt engineeringPrompt engineering didn't die, it got absorbed. The bottleneck moved from how you phrase a request to what information surrounds it. Here's what context engineering actually is, why it took over in 2026, and the discipline underneath the buzzword.
- How AI agents actually work: the loop behind the hypeEveryone is building "agents" and almost nobody explains what makes one. Not the marketing, the actual machinery: the loop, the four components, and the single property that separates an agent from a chatbot with a fancy system prompt.
Further reading
- Park et al. (2023), Generative Agents: Interactive Simulacra of Human Behavior — a memory stream with retrieval and reflection, and the clearest worked example.
- Liu et al. (2023), Lost in the Middle: How Language Models Use Long Contexts — why stuffing history into the window is not the same as the model using it.
- Packer et al. (2023), MemGPT: Towards LLMs as Operating Systems — treating the context window as managed memory.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Storing everything and retrieving badly, then concluding memory doesn't work. The failure is almost always retrieval, not storage.
- Letting the system infer sensitive facts and store them silently. Users find this unsettling, and they're right to.
- Forgetting that summaries are lossy and one-way. Whatever the summariser judged unimportant is unrecoverable.
At a glance
FieldAI Agents
Core idearemembering is retrieval, not learning
Short-termthe context window
Long-terman external store
DifficultyIntermediate
Often compared with
Agent memory vs. fine-tuning — writing facts down to look up later vs. changing the model's behaviour by training.
Where this sits
29 concepts come first. Understanding it opens up 1 more.
12Levelsteps in
29Needs firstconcepts
1Opens up0% of 310
1Areastays here
Computed from the prerequisite graph, not assigned. How this works