A tiny memory fix that stops AI agents from forgettingAI agents forget. Every time a coding assistant loses track of a debugging thread, or a data analysis agent re-ingests the same context it already processed, the team pays in latency, token costs, and brittle workflows. The fix most teams reach for — expanding the context window or adding more RAG — is increasingly expensive and still doesn't reliably work.
To address this, researchers from Mind Lab and several universities proposed delta-mem, an efficient technique that compresses the model’s historical information into a dynamically updated matrix without changing the model itself. The resulting module adds just 0.12% of the backbone model's parameters — compared to 76.40% for one leading alternative — while outperforming it on memory-heavy benchmarks. Delta-mem allows models to continuously accumulate and reuse historical data, reducing the reliance on massive context windows or complex external retrieval modules for behavioral continuity.
The long memory challenge
The conventional solution is to simply dump all the information into the model’s context window.
But as Jingdi Lei, co-author of the paper, told VentureBeat, current systems treat memory merely as a context-management problem. “Either we keep expanding the context window, or we retrieve more documents through RAG,” Lei explained. “These approaches are useful and will remain important, but they become increasingly expensive and brittle when agents need to operate over long-running, multi-step interactions, and they don't really [work] like human memory since they are more like looking up documents.”
In enterprise settings, the bottleneck is not just whether the model can access history, but whether it can reuse that history efficiently, continuously, and with low latency. Standard attention mechanisms incur a quadratic computational cost as the sequence length increases. Furthermore, expanding the context window does not guarantee the model will actually recall the information effectively. Models often suffer from context degradation or context rot as they become overwhelmed with more (and often conflicting) information, even if they support one million tokens in theory.
The researchers argue for advanced memory mechanisms that can represent historical information compactly and maintain it dynamically across interactions. Existing solutions come with heavy trade-offs and generally fall into three paradigms:
Textual memory: stores history as text injected into context — constrained by window limits and prone to information loss under compression.
Outside-channel (RAG): encodes and retrieves from external modules — adds latency, integration complexity, and potential misalignment with the backbone.
Parametric: encodes memory into model weights via adapters — static after training, can't adapt to new information during live interactions.
Inside delta-mem
To achieve a compact and dynamically updated memory, delta-mem compresses an agent’s past interactions into an “online state of associative memory” (OSAM). This state is maintained as a fixed-size matrix that preserves historical information while the underlying language model remains frozen.
For enterprise workflows, this translates directly to resolving operational bottlenecks. Lei noted that a persistent coding assistant, for example, “may need to remember project conventions, recent debugging steps, user preferences, or intermediate decisions across a workflow.” Similarly, a data analysis agent might “need to maintain task state, assumptions, and prior observations while iterating over multiple tool calls.”
Rather than repeatedly retrieving and re-inserting all relevant history for these tasks, the delta-mem matrix provides a low-overhead way to carry forward useful interaction states inside the model’s forward computation.
During generation, the system does not retrieve raw text segments to add to the prompt. Instead, the backbone LLM’s current hidden state is projected into the matrix to retrieve old memory. This operation extracts context-relevant associative memory signals from delta-mem. These signals are then transformed into numerical corrections that are applied to the computations of the model. This steers the model's reasoning at inference time without altering its internal parameters.
Following each interaction, delta-mem updates the online state using “delta-rule learning.” When new information arrives, the previous state makes a prediction about the resulting attention values. It then compares this prediction to the actual value and corrects the memory matrix based on the discrepancy.
This update mechanism relies on a “gated delta-rule.” Basically, the memory module has different knobs that control how much previous memory is kept and how much of the new memory is applied. This error correction with controlled forgetting allows the matrix to evolve over time, holding onto stable historical associations without being derailed by short-term noise.
The researchers explored three strategies for determining when and how the matrix updates:
Token-state write captures fine-grained changes but is vulnerable to short-term noise.
Sequence-state write averages tokens within a message segment, smoothing updates at the cost of some localized detail.
Multi-state write decomposes memory into sub-states for different information types like facts or task progress.
Delta-mem in action
The researchers evaluated delta-mem across three LLM backbones: Qwen3-8B, Qwen3-4B-Instruct, and SmolLM3-3B. They configured the framework with a compact 8x8 matrix. The system was tested on general capability benchmarks, including HotpotQA, GPQA-Diamond, and IFEval. It was also evaluated on memory-heavy tasks such as LoCoMo, which tests long-term conversational memory, and Memory Agent Bench, which assesses retention, retrieval, selective forgetting, and test-time learning over extended interactions.
The framework was compared against representative models from the three existing memory paradigms: textual memory baselines (e.g., BM25 RAG, LLMLingua-2, and MemoryBank), parametric systems (Context2LoRA and MemGen), and the outside-channel approach MLP Memory.
Across the board, delta-mem outperformed the baselines, according to the researchers. On the Qwen3-4B-Instruct backbone, the token-state write variant achieved an average score of 51.66%, easily surpassing the frozen vanilla backbone at 46.79% and the strongest baseline, Context2LoRA, at 44.90%. On the memory-heavy Memory Agent Bench, the average score jumped from 29.54% to 38.85%. Performance on the specific test-time learning subtask nearly doubled from 26.14 to 50.50.
However, the most compelling takeaways are the system's operational efficiency. The researchers tested the framework in a no-context setting where the historical text was entirely removed from the context. Even without explicit text replay, delta-mem successfully recovered context-relevant evidence in multi-hop tasks. The researchers argue that the model remembers past interactions without needing to ingest massive amounts of prompt tokens.
The framework also adds only 4.87 million trainable parameters, representing just 0.12% of the Qwen3-4B-Instruct backbone. By comparison, the MLP Memory baseline required 3 billion parameters, scaling up to 76.40% of the backbone's size while delivering inferior results. When prompt lengths scaled up to 32,000 tokens during inference tests, the framework maintained almost the exact same GPU memory footprint as a standard, unmodified model. It sidesteps the heavy memory bloat that affects other advanced memory systems like MemGen and MLP Memory.
Different update strategies proved beneficial depending on the underlying model capacity. The sequence-state write strategy was the most effective for stronger backbones like Qwen3-8B. These more capable models use the segment-level writing to smooth out updates and mitigate token-level noise. Conversely, the multi-state write strategy drove massive performance leaps for smaller backbones like SmolLM3-3B. For these lower-capacity models, separating memory into multiple states proved critical to minimizing information interference.
Implementing delta-mem in the enterprise stack
The researchers have released the code for delta-mem on GitHub and the weights for their trained adapters on Hugging Face. For AI engineering teams looking to integrate this framework into their existing inference stack, the process requires minimal computing resources.
“In practice, an engineering team would start from an existing instruction-tuned backbone, attach the Delta-Mem adapter modules to selected attention layers, train only the adapter parameters on domain-relevant multi-turn or long-context data... and then run inference with the memory state updated online during interaction,” Lei said. Crucially, teams do not need a massive pretraining corpus. The training data only needs to reflect the target memory behavior, such as multi-turn dialogues, agent traces, or domain workflows where earlier information must influence later decisions.
While compressing interaction history into a fixed-size mathematical matrix creates immense efficiency, it does come with trade-offs. Delta-mem is not a lossless replacement for explicit text logs or document retrieval. Because different pieces of information compete inside the same limited state, there is a risk of memory blending.
“Delta-Mem is useful when the system needs fast, online, continuously updated behavioral state,” Lei said. “RAG is better when the system needs exact factual recall, citation, compliance, auditability, or access to a large external knowledge base.” Remembering a user’s working style or a multi-step reasoning trajectory is a perfect fit for delta-mem, while retrieving a legal contract or a medical guideline should remain in a vector database.
This means the most realistic enterprise architecture moving forward is a hybrid approach. Delta-mem acts as a lightweight internal working memory, reducing the need to retrieve or replay everything all the time, while RAG serves as the explicit, high-capacity memory layer.
“Looking ahead, I do not think vector databases will become obsolete,” Lei said. “Instead, I expect enterprise AI stacks to become more layered. We will likely see short-term working memory inside the model, longer-term explicit memory in retrieval systems, and policy or audit layers that decide what should be stored, retrieved, forgotten, or exposed to the user.”
Explore how xMemory tames context bloat for persistent AI agentsStandard RAG pipelines break when enterprises try to use them for long-term, multi-session LLM agent deployments. This is a critical limitation as demand for persistent AI assistants grows.
xMemory, a new technique developed by researchers at King’s College London and The Alan Turing Institute, solves this by organizing conversations into a searchable hierarchy of semantic themes.
Experiments show that xMemory improves answer quality and long-range reasoning across various LLMs while cutting inference costs. According to the researchers, it drops token usage from over 9,000 to roughly 4,700 tokens per query compared to existing systems on some tasks.
For real-world enterprise applications like personalized AI assistants and multi-session decision support tools, this means organizations can deploy more reliable, context-aware agents capable of maintaining coherent long-term memory without blowing up computational expenses.
RAG wasn't built for this
In many enterprise LLM applications, a critical expectation is that these systems will maintain coherence and personalization across long, multi-session interactions. To support this long-term reasoning, one common approach is to use standard RAG: store past dialogues and events, retrieve a fixed number of top matches based on embedding similarity, and concatenate them into a context window to generate answers.
However, traditional RAG is built for large databases where the retrieved documents are highly diverse. The main challenge is filtering out entirely irrelevant information. An AI agent's memory, by contrast, is a bounded and continuous stream of conversation, meaning the stored data chunks are highly correlated and frequently contain near-duplicates.
To understand why simply increasing the context window doesn’t work, consider how standard RAG handles a concept like citrus fruit.
Imagine a user has had many conversations saying things like “I love oranges,” “I like mandarins,” and separately, other conversations about what counts as a citrus fruit. Traditional RAG may treat all of these as semantically close and keep retrieving similar “citrus-like” snippets.
“If retrieval collapses onto whichever cluster is densest in embedding space, the agent may get many highly similar passages about preference, while missing the category facts needed to answer the actual query,” Lin Gui, co-author of the paper, told VentureBeat.
A common fix for engineering teams is to apply post-retrieval pruning or compression to filter out the noise. These methods assume that the retrieved passages are highly diverse and that irrelevant noise patterns can be cleanly separated from useful facts.
This approach falls short in conversational agent memory because human dialogue is “temporally entangled,” the researchers write. Conversational memory relies heavily on co-references, ellipsis, and strict timeline dependencies. Because of this interconnectedness, traditional pruning tools often accidentally delete important bits of a conversation, leaving the AI without vital context needed to reason accurately.
Why the fix most teams reach for makes things worse
To overcome these limitations, the researchers propose a shift in how agent memory is built and searched, which they describe as “decoupling to aggregation.”
Instead of matching user queries directly against raw, overlapping chat logs, the system organizes the conversation into a hierarchical structure. First it decouples the conversation stream into distinct, standalone semantic components. These individual facts are then aggregated into a higher-level structural hierarchy of themes.
When the AI needs to recall information, it searches top-down through the hierarchy, going from themes to semantics and finally to raw snippets. This approach avoids redundancy. If two dialogue snippets have similar embeddings, the system is unlikely to retrieve them together if they have been assigned to different semantic components.
For this architecture to succeed, it must balance two vital structural properties. The semantic components must be sufficiently differentiated to prevent the AI from retrieving redundant data. At the same time, the higher-level aggregations must remain semantically faithful to the original context to ensure the model can craft accurate answers.
A four-level hierarchy that shrinks the context window
The researchers developed xMemory, a framework that combines structured memory management with an adaptive, top-down search strategy.
xMemory continuously organizes the raw stream of conversation into a structured, four-level hierarchy. At the base are the raw messages, which are first summarized into contiguous blocks called “episodes.” From these episodes, the system distills reusable facts as semantics that disentangle the core, long-term knowledge from repetitive chat logs. Finally, related semantics are grouped together into high-level themes to make them easily searchable.
xMemory uses a special objective function to constantly optimize how it groups these items. This prevents categories from becoming too bloated, which slows down search, or too fragmented, which weakens the model’s ability to aggregate evidence and answer questions.
When it receives a prompt, xMemory performs a top-down retrieval across this hierarchy. It starts at the theme and semantic levels, selecting a diverse, compact set of relevant facts. This is crucial for real-world applications where user queries often require gathering descriptions across multiple topics or chaining connected facts together for complex, multi-hop reasoning.
Once it has this high-level skeleton of facts, the system controls redundancy through what the researchers call "Uncertainty Gating." It only drills down to pull the finer, raw evidence at the episode or message level if that specific detail measurably decreases the model’s uncertainty.
“Semantic similarity is a candidate-generation signal; uncertainty is a decision signal,” Gui said. “Similarity tells you what is nearby. Uncertainty tells you what is actually worth paying for in the prompt budget.” It stops expanding when it detects that adding more detail no longer helps answer the question.
What are the alternatives?
Existing agent memory systems generally fall into two structural categories: flat designs and structured designs. Both suffer from fundamental limitations.
Flat approaches such as MemGPT log raw dialogue or minimally processed traces. This captures the conversation but accumulates massive redundancy and increases retrieval costs as the history grows longer.
Structured systems such as A-MEM and MemoryOS try to solve this by organizing memories into hierarchies or graphs. However, they still rely on raw or minimally processed text as their primary retrieval unit, often pulling in extensive, bloated contexts. These systems also depend heavily on LLM-generated memory records that have strict schema constraints. If the AI deviates slightly in its formatting, it can cause memory failure.
xMemory addresses these limitations through its optimized memory construction scheme, hierarchical retrieval, and dynamic restructuring of its memory as it grows larger.
When to use xMemory
For enterprise architects, knowing when to adopt this architecture over standard RAG is critical. According to Gui, “xMemory is most compelling where the system needs to stay coherent across weeks or months of interaction.”
Customer support agents, for instance, benefit greatly from this approach because they must remember stable user preferences, past incidents, and account-specific context without repeatedly pulling up near-duplicate support tickets. Personalized coaching is another ideal use case, requiring the AI to separate enduring user traits from episodic, day-to-day details.
Conversely, if an enterprise is building an AI to chat with a repository of files, such as policy manuals or technical documentation, “a simpler RAG stack is still the better engineering choice,” Gui said. In those static, document-centric scenarios, the corpus is diverse enough that standard nearest-neighbor retrieval works perfectly well without the operational overhead of hierarchical memory.
The write tax is worth it
xMemory cuts the latency bottleneck associated with the LLM's final answer generation. In standard RAG systems, the LLM is forced to read and process a bloated context window full of redundant dialogue. Because xMemory's precise, top-down retrieval builds a much smaller, highly targeted context window, the reader LLM spends far less compute time analyzing the prompt and generating the final output.
In their experiments on long-context tasks, both open and closed models equipped with xMemory outperformed other baselines, using considerably fewer tokens while increasing task accuracy.
However, this efficient retrieval comes with an upfront cost. For an enterprise deployment, the catch with xMemory is that it trades a massive read tax for an upfront write tax. While it ultimately makes answering user queries faster and cheaper, maintaining its sophisticated architecture requires substantial background processing.
Unlike standard RAG pipelines, which cheaply dump raw text embeddings into a database, xMemory must execute multiple auxiliary LLM calls to detect conversation boundaries, summarize episodes, extract long-term semantic facts, and synthesize overarching themes.
Furthermore, xMemory’s restructuring process adds additional computational requirements as the AI must curate, link, and update its own internal filing system. To manage this operational complexity in production, teams can execute this heavy restructuring asynchronously or in micro-batches rather than synchronously blocking the user's query.
For developers eager to prototype, the xMemory code is publicly available on GitHub under an MIT license, making it viable for commercial uses. If you are trying to implement this in existing orchestration tools like LangChain, Gui advises focusing on the core innovation first: “The most important thing to build first is not a fancier retriever prompt. It is the memory decomposition layer. If you get only one thing right first, make it the indexing and decomposition logic.”
Retrieval isn't the last bottleneck
While xMemory offers a powerful solution to today's context-window limitations, it clears the path for the next generation of challenges in agentic workflows. As AI agents collaborate over longer horizons, simply finding the right information won't be enough.
“Retrieval is a bottleneck, but once retrieval improves, these systems quickly run into lifecycle management and memory governance as the next bottlenecks,” Gui said. Navigating how data should decay, handling user privacy, and maintaining shared memory across multiple agents is exactly “where I expect a lot of the next wave of work to happen,” he said.
