The A-MEM paper proposes memory notes that can gain links, tags, context, and embeddings as an agent encounters new information.
2
A-MEM uses an LLM to examine new memories alongside nearby memories, then update the surrounding metadata and connections.
3
Nehil Jain reported that A-MEM performed especially well on multi-hop conversations, while the evaluation left open questions about latency, runaway updates, and generalisation beyond conversational data.
Summary
The paper discussed was A-MEM: Agentic Memory for LLM Agents. Matt Squire introduced its motivation through a DevOps agent that remembers infrastructure failures, approved fixes, and the lessons learned from operating production systems. The paper's authors propose a Zettelkasten-inspired memory system made from small notes containing original content, generated context, keywords, tags, timestamps, embeddings, and links to related notes. Adam Becker walked through the implementation: an LLM enriches each new note, vector search finds nearby memories, and another LLM decides whether to create links or update existing metadata. Nehil Jain reviewed the evaluation on the LOCOMO dataset and said the strongest results appeared on multi-hop questions, with linking and evolution both contributing to performance. The discussion questioned the tag ontology, the choice of nearest-neighbour count, the cost and timing of updates, the absence of clear rollback behaviour, and whether conversational benchmarks say enough about production agents.
A production agent needs memory of actions and approved fixes
Matt Squire motivated the paper with a DevOps or site reliability agent responsible for services, databases, replicas, and systems spread across cloud environments. If a service crashes, the agent could inspect logs, examine code, suggest a fix, and send it to a human for approval. The useful memory includes the original failure, the proposed solution, the action taken, and the human feedback. On a later incident, the agent could use that record instead of treating the event as new. Matt used an example where a service failed because its node did not have enough memory provisioned.
A-MEM avoids a fixed schema for long-lived agent knowledge
Matt said the paper's authors criticise existing memory systems, including graph databases, for relying on predefined schemas. Their proposed system aims to let an agent build knowledge in complex and changing environments without imposing a rigid structure. A memory can contain text, tags, context, embeddings, and links. The system can also update existing memories when new information changes how they relate to one another. Matt described the intended result as a flexible network of small facts and connections that grows as the agent interacts with its environment.
Zettelkasten supplies the paper's model for atomic notes and references
Adam Becker explained the Zettelkasten method as an influence on A-MEM. The method organises knowledge into small, discrete notes, gives them identifiers, adds tags, and connects related notes. A note can have a nested identifier such as 1/1/1, while another note can reference it through its identifier. Adam described the paper's innovation as applying this approach to agent memory: a new memory should be linked to relevant existing memories, and the wider memory structure should evolve when that new note changes the surrounding context.
Each new memory is enriched, embedded, and compared with nearby memories
Adam used a short conversation about taking up photography as an example. The original content is passed to an AI system that generates a contextual description, tags such as hobby and photography, and keywords such as photography and scenery. The textual fields are concatenated and embedded. Vector search then finds nearby memories. An LLM examines those candidates more closely and decides whether links should be created and whether their context, tags, or keywords should change. Adam said the original content remains the seed of the memory while the surrounding metadata can evolve.
The system combines rough retrieval with LLM-based memory evolution
Adam showed the operational flow for a request about implementing a custom cache system. The agent creates a note, enriches it, retrieves the nearest memories from a vector database, and examines those memories for possible links or updates. Retrieval can use cosine similarity, keyword matching, or a weighted combination of the two. The implementation includes a memory note class, a retriever choice, an evolution decision, and periodic consolidation into the database rather than writing every change immediately. Adam said the open-source code contains these pieces, although he did not find it very clean.
A-MEM's reported advantage is strongest on multi-hop conversations
Nehil Jain reviewed the LOCOMO evaluation, which contains long conversations with many sessions and tokens. The comparison included putting the available conversation directly into the model, ReadAgent, MemoryBank, and MemGPT. Nehil said A-MEM did not win every category, especially open-domain and adversarial settings, but it performed especially well on multi-hop questions that require information from multiple sessions. He also said it was cheaper in that evaluation. The metrics included F1, BLEU, ROUGE, semantic distance, and the cost of placing memory into context.
Linking and evolution both contributed in the paper's ablation study
Nehil described an experiment that removed either linking or evolution from the system. The best result came when both were enabled, which suggested that the two mechanisms each added to performance in the reported study. He also discussed tests of different values of K, the number of nearest neighbours considered during memory processing. The paper used K equal to 30 in one setting, but Nehil questioned whether the best value should depend on the model and whether the reported experiments were broad enough to support a general choice.
The evaluation leaves practical questions about safety, cost, and scope
Nehil questioned how the system handles runaway evolution, where repeated updates could create cascading changes that are difficult to control or undo. He also noted that LOCOMO focuses on conversational memory and does not test the code-related infrastructure example that motivated the discussion. Other open questions included how the system works with structured databases, how much latency the retrieval and evolution steps add, and whether the evaluation covers general-purpose tasks. The group also discussed whether keeping the history of memory changes would help with entity resolution and later validation.
"Now that's that kind of system would be all the more interesting if it can learn about the infrastructure that it's managing as it goes along."Matt Squire07:37
Who should watch
You are building an agent that needs to remember incidents, user interactions, or approved actions across long-running sessions.
You are deciding between vector retrieval, graph-style links, and LLM-generated metadata for an agent memory system.
You need to assess a memory paper beyond its headline results and want to examine benchmark choice, update safety, cost, and production latency.
