Still calling an LLM for every single memory operation your agent makes? This paper splits that work between a "fast brain" and a "slow brain" instead. Title: Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents URL: Inspired by dual-process cognition (System One vs. System Two), Jev-Mem handles most memory operations with lightweight structured decisions and reserves the LLM for genuinely complex reasoning. Three highlights stand out. 🧠 Typed System-One control Frequent operations like typing, relation judgment, and query routing run through a lightweight interface that returns probabilities and labels instead of free-form text — the same control layer governs both memory construction and retrieval. 🕸️ A four-relation memory graph Memory is organized across semantic, temporal, causal, and entity relations, with retrieval budget allocated to whichever views matter most for a given query — avoiding wasted traversal. 📊 Accuracy and speed improve together On the LoCoMo benchmark, Jev-Mem scores 0.777 overall, an 11% gain over the best baseline, while building memory 6.6x faster (158 seconds) and answering queries 36.7% faster (0.93 seconds). What stands out to me is that separating control from reasoning improved accuracy and efficiency at the same time, not one at the expense of the other. #AIAgents# #MemoryArchitecture#

This might be the first open-source memory architecture to cover associative recall. The rest stop at similarity search. (T-Mem, EMNLP 2026) Most memory systems retrieve by similarity. T-Mem predicts when a memory will matter, before you ask. → write: memories ship with a trigger at store time, basically "when will this come up again?" → read: at query time, triggers pull in memories by situation, even with zero keyword overlap. → architecture layer. plugs into any LLM without fine-tuning or model changes. On LoCoMo-Plus (strips keyword overlap, tests pure associative recall), mainstream systems drop 28 to 50 points. T-Mem drops 5.45. paper: code: (MIT)

Introducing MemCurrent, a persistent memory architecture designed to help production agents stop starting from zero. MemCurrent combines episodic memory with derived memory across three access patterns: • Plugins to log and load context automatically • Tools the agent calls on demand • An offline “dreaming” pipeline that consolidates memory in the background Agents carry forward user preferences, prior corrections, metric definitions and successful workflows instead of rediscovering them every session. On our internal benchmark built from real enterprise data, MemCurrent triples the agent's pass rate while using 31% fewer tokens and running 26% faster. It's built into Agent Bricks' Memory offering and already powers memory in Genie. Read the research:

# Decision Points for Embedding AI Agents in Enterprise Systems # Centralized vs Federated Memory — Where to Store Agent Memory 🎯 The Hook Should your agent's memory — user profiles, conversation history, learned data — live in a single central store or stay distributed across each data source? This is not merely a technical choice. It directly impacts data sovereignty, cross-border regulatory compliance, and personalization quality. For globally operating enterprises, GDPR and China's Data Security Law can fundamentally reshape the architecture 🔑 📋 Overview Centralized memory has all agents referencing a single central store, enabling consistent personalization across multiple channels — using Slack conversation history in a Salesforce agent, for example. Federated memory keeps data distributed across source systems, with agents fetching only when needed, making compliance with data sovereignty and cross-border transfer regulations straightforward. Both have strengths and limitations. Most enterprises settle on a hybrid: metadata centralized, detailed data federated 📊 🔍 Decision Points The most critical axis is cross-border data transfer regulation. If the target data falls under GDPR, China's Data Security Law, or similar cross-border transfer regulations, federated memory is the default. When data cannot be moved from its source system, centralization is off the table. If consistent personalization across multiple systems is required, consider centralization. This applies to cases requiring strong consistency — like ensuring all agents immediately reflect a customer's cancellation intent. If access permission models across data sources cannot be unified, use federation to respect each source's permissions. Consolidation risks violating the principle of least privilege. If data volume and update frequency make synchronization costs to a central store prohibitive, consider federation with a cache layer ⚡ 💡 Key Details Three hybrid approaches to understand: Metadata centralized, detail data federated — Lightweight metadata (user identifiers, last access dates, preference summaries) lives centrally. Detailed data (conversation log bodies, attachments) stays at each source. This is the most common pattern, balancing central convenience with federated data sovereignty compliance. Virtual centralization via cache layer — Fetch results from each source in real time, integrate them in a short-lived cache, and build a pseudo-centralized view per request. This avoids permanent data replication, making GDPR right-to-erasure compliance easier. Context broker — Place a broker that collects and combines needed data from each source at request time. Provides cross-cutting context without a central store, but latency and source-failure fallback design become challenges 🔄 ⚖️ Trade-offs Replicating all data to a central store inflates storage costs and introduces sync delays with source data. Updating customer information in Salesforce while the centralized memory still shows the old version — this inconsistency fundamentally undermines agent trustworthiness. Right-to-erasure compliance also becomes complex: if deletion of centrally copied data does not synchronize with source deletion, it becomes a GDPR risk 😩 On the other hand, forcing cross-cutting queries on a fully federated architecture means multi-hop real-time queries across sources, pushing latency past 5 seconds and degrading user experience. Source outages cause the agent to return errors. Fallback design combining cached responses with freshness warnings is essential. The most dangerous failure is missing data classification entirely. Not classifying which data can be centralized and which cannot cross borders before build — then discovering compliance violations in production — is more common than you might think ⚠️ 🛠️ Use Cases Global customer engagement agent (Salesforce/Zendesk integration): EU customer data stays in the EU-region Zendesk, Japanese customer data in the Japan-region Salesforce — federated by regulation. Only metadata (customer ID, last interaction date, preference summary) is aggregated into a central CDP, giving all agents a consistent customer understanding. Detailed data is fetched from each source on demand 📚 Internal knowledge agent (ServiceNow/Workday/Confluence integration): ServiceNow incident history is under IT department governance, Workday HR data under HR department governance, each source controlling its own access permissions. A context broker collects and combines needed data from each source at request time to generate cross-cutting answers 🏢 Customer 360 agent (Salesforce CDP integration): A unified company-wide customer profile store at the center, with all agents sharing the same customer understanding. State changes like cancellation intent or critical complaints are reflected immediately with strong consistency. Best suited for domestic operations where data fits within a single region 🔧 Practical tip: Start by establishing a data classification policy. Categorize data into three buckets — "centralizable," "cannot cross borders," and "access-restricted" — and select your memory architecture based on that classification. Start the design from compliance requirements, not technology selection 💪 #AIAgents# #EnterpriseArchitecture#

Agents shouldn't have to relearn the same facts every session. Agent Memory on Agent Bricks combines session history with distilled long-term memories. A background pipeline extracts durable facts, removes duplicates, and replaces information contradicted by newer conversations. Lakebase provides low-latency access. In our internal benchmarks of enterprise data tasks, this memory architecture tripled agent pass rates while using 31% fewer tokens compared to agents without memory.

At SEMICON Taiwan’s imec Technology Forum, memory architecture and 3D integration emerged as closely connected parts of this challenge. Gouri Sankar Kar examined next-generation memory hierarchies for edge, physical, and agentic AI, while Zsolt Tőkei addressed the interconnect and integration technologies needed for future computing systems. A common direction was clear: select memory according to how data is used, then connect it vertically at the density and proximity the workload requires.

I read LS Securities’ report, “Memory Shortages Even in 2030.” (I love this title name) First, thanks to @PhotonCap for sharing it with me. TL;DR: KV cache offload and HBM optimization do not necessarily mean lower overall memory demand. As AI advances, memory is likely to expand beyond HBM into more layers such as local DRAM, CXL, and storage, while becoming increasingly specialized for different workloads. On the supply side, even as fabs and wafer capacity continue to expand, improving bits per wafer through DRAM scaling is becoming increasingly difficult. At the same time, a meaningful portion of incremental DRAM bits may be allocated to HBM, so wafer growth should not be interpreted directly as conventional DRAM supply growth. On the demand side, memory content per system is becoming more important than CPU or GPU unit growth. Even if server shipments do not increase dramatically, more memory channels, higher-density DIMMs, and the expanding memory footprint of AI workloads can continue to drive memory demand per system higher. This becomes particularly important with Agentic AI. As context windows, KV cache, and agent state grow, the memory problem expands from bandwidth alone to capacity. Not everything can reside in HBM, which naturally leads toward a broader memory hierarchy: HBM → Local DRAM → CXL → Storage This also connects with the custom memory thesis I have discussed for some time. Rather than relying on one general-purpose memory architecture, I expect AI systems to increasingly optimize bandwidth, capacity, and latency for specific workloads. This is why I continue to pay attention to architectures such as Custom HBM, Memory-on-Logic, zHBM, and HBF. And as memory becomes more distributed across these tiers, the next bottleneck increasingly becomes data movement. Accessing a larger memory pool is only useful if data can move between these tiers with sufficient bandwidth, low latency, and low power. This is also why I continue to view CXL, fabric, CPO, and optics as closely connected to the evolution of memory architecture. Memory optimization does not mean less memory. It means memory moves. AI systems are becoming more memory-heavy, memory itself is becoming more tiered and specialized, and the fabrics and optics connecting those memory pools should become increasingly important as well.