Releases·Tool Release
Jev-Mem: System-One-controlled agentic memory cuts LoCoMo latency and boosts answer quality
Dongming Jiang, Yi Li, and Bingzhe Li released Jev-Mem, an agentic memory system that separates fast memory-management decisions (a lightweight System-One controller) from deep answer synthesis (a System-Two language model). On LoCoMo with GPT-4o-mini it reports higher answer quality, faster memory construction, and lower query latency than baselines. The code, arXiv paper, and Hugging Face demo are public, making it directly usable for developers building long-running agents.
90 engagement1 sourceSun, Sep 20, 2026, 05:27 AM
AI Overview
- Paper by Dongming Jiang, Yi Li, and Bingzhe Li separates a lightweight System-One memory controller from a System-Two LLM answer synthesizer
- On LoCoMo with GPT-4o-mini, Jev-Mem reports 11.0% higher answer quality, 6.6× faster memory construction, and 36.7% lower query latency than baselines
- System-One controller organizes retrieval across semantic, temporal, causal, and entity relations; code and demo are on GitHub and Hugging Face Spaces
Related
Developments·Paper
Just-In-Time Agent Memory with Runtime Agentic Research
A new paper introduces Just-In-Time Agent Memory (JAM), a trainable framework that constructs query-conditioned context at runtime instead of pre-building memory ahead of time. A Memorizer keeps raw histories in a hierarchical page-store while a Researcher iteratively retrieves and integrates evidence per request. JAM outperforms Ahead-of-Time memory systems on agent-memory and long-context benchmarks while staying more efficient than prior trained agentic memory approaches.
A new paper introduces Just-In-Time Agent Memory (JAM), a trainable framework that constructs query-conditioned context at runtime instead of pre-building memory ahead of time. A Memorizer keeps raw histories in a hierarchical page-store while a Researcher iteratively retrieves and integrates evidence per request. JAM outperforms Ahead-of-Time memory systems on agent-memory and long-context benchmarks while staying more efficient than prior trained agentic memory approaches.
0 engagement1 sourcearxivMon, Sep 28, 2026, 06:01 AM
Releases·Tool Release
SodaMem releases evidence-grounded temporal memory for AI agents
SodaMem, an open-source agentic memory infrastructure, has been released on GitHub. It provides evidence-grounded temporal memory where each memory names the turn it originated from and tracks when it became invalid. The project reports strong benchmark results on LongMemEval and LoCoMo.
SodaMem, an open-source agentic memory infrastructure, has been released on GitHub. It provides evidence-grounded temporal memory where each memory names the turn it originated from and tracks when it became invalid. The project reports strong benchmark results on LongMemEval and LoCoMo.
24 engagement1 sourcegithubFri, Aug 7, 2026, 09:46 AM
Developments·Paper
PSD distills memory-construction from GPT-4.1-mini into Qwen3 SLMs
A new arXiv paper introduces Pseudo Self-Distillation (PSD), a framework letting small language models build hierarchical agent memory by distilling a black-box oracle's behavior without access to its logits or hidden states. On LoCoMo, PSD-trained Qwen3-0.6B/1.7B/4B match or exceed GPT-4.1-mini on retrieval at a fraction of the cost, with capability transferring out-of-distribution to LongMemEval.
A new arXiv paper introduces Pseudo Self-Distillation (PSD), a framework letting small language models build hierarchical agent memory by distilling a black-box oracle's behavior without access to its logits or hidden states. On LoCoMo, PSD-trained Qwen3-0.6B/1.7B/4B match or exceed GPT-4.1-mini on retrieval at a fraction of the cost, with capability transferring out-of-distribution to LongMemEval.
0 engagement1 sourcearxivSun, Sep 20, 2026, 08:24 AM
Developments·Paper
EGMEMORY: Evidence-Grounded Memory for Multi-Actor Conversations Hits 77.9% on EverMemBench
Two arXiv papers (plus a GitHub release) tackle long-horizon memory in multi-party dialogue, where systems must track who said what, group versus personal scope, and how states change over time. SpeakerMem-R1 uses a dual-track memory with an RL-trained writer (SpeakerLevenshtein, speaker-conditioned GRPO), while EGMEMORY models memory as a searchable state machine with a propose-verify-commit protocol and no memory-specific training. Both report strong results on GroupMemBench, EverMemBench, and LoCoMo, though their headline numbers differ, relevant to practitioners building agent memory systems.
Two arXiv papers (plus a GitHub release) tackle long-horizon memory in multi-party dialogue, where systems must track who said what, group versus personal scope, and how states change over time. SpeakerMem-R1 uses a dual-track memory with an RL-trained writer (SpeakerLevenshtein, speaker-conditioned GRPO), while EGMEMORY models memory as a searchable state machine with a propose-verify-commit protocol and no memory-specific training. Both report strong results on GroupMemBench, EverMemBench, and LoCoMo, though their headline numbers differ, relevant to practitioners building agent memory systems.
03073 engagement3 sourcesarxiv, githubSun, Sep 20, 2026, 08:51 AM
Developments·Paper
MERA: RL-trained 0.6B planner retrieves missing evidence for long-term memory QA
Researchers introduce MERA (Missing-Evidence Retrieval Augmentation), a method for long-term memory QA where evidence is scattered across distant conversation turns. It separates globally searchable memory from a question-specific evidence state and trains a lightweight 0.6B planner via reinforcement learning, rewarded for queries that recover previously missing evidence. The approach improves answer accuracy on Qwen3-30B and GPT-4o-mini backbones, offering practitioners a compute-cheap planner that outperforms much larger untrained planners.
Researchers introduce MERA (Missing-Evidence Retrieval Augmentation), a method for long-term memory QA where evidence is scattered across distant conversation turns. It separates globally searchable memory from a question-specific evidence state and trains a lightweight 0.6B planner via reinforcement learning, rewarded for queries that recover previously missing evidence. The approach improves answer accuracy on Qwen3-30B and GPT-4o-mini backbones, offering practitioners a compute-cheap planner that outperforms much larger untrained planners.
0 engagement1 sourcearxivTue, Sep 29, 2026, 01:02 PM