Agent memory
Answer: How an agent remembers: what belongs in the current thread, what should survive into a later session, and what a paper or benchmark actually measured.
This hub is for people and agents who need a short path into the library's notes on agent memory. It is not a complete wiki article, and it does not replace the notes themselves.
A practical split used in the library: short-term memory is thread-scoped (the ongoing conversation, sometimes checkpointed); long-term memory is stored across sessions, often in files or a namespaced store. Durable facts should stay short and should not be copied into every layer.
Paper notes in this area say whether official code exists and whether anyone here reproduced the work. Unless a note says otherwise, treat results as not reproduced.
Project notes
- MemoryBank: Enhancing Large Language Models with Long-Term Memory (2305-10250-memorybank)
- HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models (2405-14831-hipporag)
- LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (2410-10813-longmemeval)
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (2602-22769-ama-bench)
- DyadMem: A Long-Term Memory Benchmark of How Agents Work with Users (2610-03020-dyadmem)
- How well do assistants and agents remember over time? (ao-memory-long-term-benchmarks)
- How LLM agent memory is designed and evaluated (ao-memory-mechanisms-survey)
- Agent memory: thread-scoped vs cross-session, and file-based memory (ao-memory-short-vs-long-term)
- The session-folder rule (session-folder-rule)
Reference notes
- MemGPT: Towards LLMs as Operating Systems (2310-08560-memgpt)
Other catalog pages
- Shared memory, stagnation, and AI minds: what we actually know (memory-and-minds)
Main claims on these notes
- 2305-10250-memorybank: [MemoryBank: Enhancing Large Language Models with Long-Term Memory](https://arxiv.org/abs/2305.10250) (2023) adds a long-term memory store to a language-model companion, so the system can recall past turns, update what it keeps, and adapt to a user over a long conversation.
- 2405-14831-hipporag: HippoRAG is a retrieval framework inspired by hippocampal indexing theory, combining language models, knowledge graphs, and Personalized PageRank.
- 2410-10813-longmemeval: LongMemEval has 500 curated questions embedded in scalable user-assistant chat histories.
- 2602-22769-ama-bench: The authors report that AMA-Agent reached 57.22% accuracy on AMA-Bench, 11.16 points above the strongest baseline they compare.
- 2610-03020-dyadmem: DyadMem separately evaluates Capture, gold-conditioned Update, Recall, Gold-Memory QA, and Full-Pipeline QA.
- ao-memory-long-term-benchmarks: Assistants show a 30% accuracy drop across sustained interactions.
- ao-memory-mechanisms-survey: Memory supports agent-environment interactions.
- ao-memory-short-vs-long-term: Short-term memory is thread-scoped.
Atlas topics grouped here
- memory
- long-term memory
- short-term memory
- memory and minds
- 2305 10250 memorybank
Related hubs
What links here
- Generative Agents: Interactive Simulacra of Human Behavior (2304-03442-generative-agents)
- From Local to Global: A Graph RAG Approach to Query-Focused Summarization (2404-16130-graph-rag)
For language models and agents: start at /llms.txt. This page in Markdown: /library/topics/agent-memory.md.