Yesterday's mistake: AI agents will make it again today. Their weights are frozen and their context is finite. Fine-tuning risks catastrophic forgetting, and most memory systems are just places where the past can be looked up. This paper changes the question: judge memory not by what it retrieves, but by whether it improves the next action. It proposes MCCL, a framework for continual learning through memory alone, with no updates to the foundation model. It organizes memory along three axes (where it lives, what form it takes, how it is managed) and describes two loops: an inner loop that writes experience, and an outer loop that improves the management policy itself. Evaluation is split into four gates (preserve, access, use, effect) to locate where a failure happens. It is a research agenda without experiments, but a useful lens for rethinking agent memory design. Paper: Memory-Centric Continual Learning for AI Agents #AIAgents# #ContinualLearning#

Can an agent that "improves itself" really learn new things without forgetting the old? Title: Are Self-Evolving Agents Continual Learners? URL: ❓ What did they study? Self-evolving agents that keep model weights frozen and rewrite the harness (prompts, memory, skills): ReasoningBank, GEPA, and Meta-Harness, tested on task streams where topic or workflow shifts midway. 💡 What happened? Performance rose in most settings (GEPA gained +3.7pp at the end on average). But the more adaptable a method was, the less stable it was. In all 12 method-stream pairs, the most plastic method was less stable than the most conservative one. ❓ Why does it forget? Methods that rewrite a shared prompt or workspace make global edits that interfere with unseen tasks. GEPA condensed seven BIRD sections into one and BIRD-F1's gain fell back to 0.0. ❓ Is memory retrieval the fix? It localizes updates, but not perfectly. Before a group's block it always retrieves a foreign memory, and for two similar workflows it pulled the wrong memory in 52% of episodes, cutting accuracy by 8.0pp. #AIAgents# #ContinualLearning#

TL;DR: Cortex bakes the gains an LLM gets from context into its weights, so they persist after the context is removed. After 50 sequential factual edits it recovers 88.9% of the gain while keeping 99.8% of original performance. Title: Cortex: Consolidating Context into Parametric Memory URL: Key points ・🧠 Inspired by hippocampus-to-neocortex consolidation; unifies knowledge editing, skill learning, behavior steering, and suppression as one problem ・🔮 The model generates queries anticipating how the context will be used and learns from them ・🧩 Updates only the FFN: low-rank updates (Synaptic) or new added channels (Structural), both mergeable into the weights ・👩‍🏫 A frozen teacher that sees the context distills its outputs and hidden states into a student that doesn't ・📊 Tested on 4B-120B dense and MoE models; batch editing hits 80-91% CR vs. 36-46% for the strongest baselines ・🤖 On ALFWorld, the consolidated model even beats supplying the context directly ・⚠️ Knowledge suppression still harms neighboring knowledge 💡 A practical path beyond RAG and prompting: folding experience into the weights. #LLM# #ContinualLearning#

🧠 Once an agent writes down a lesson that turns out to be wrong, it just keeps believing it forever. Here's a paper that tackles that quiet but serious problem. Long-running LLM agents write down "reflections" from past experience and read them back later to decide what to do next. But existing harnesses only let information flow forward: once a lesson is written, nothing ever revisits it, even after the world has moved on. A database schema changes and the agent keeps assuming the old columns. A single failure gets over-generalized into a rule the agent follows forever. These stale memories quietly keep misleading the agent. That's the problem ReHarness, proposed in "Recurrent Harness for Agentic Continual Learning," sets out to fix. On top of the usual forward write and read, it adds a "backward revision" step: using feedback from the current experience to rewrite the reflections that were just retrieved, while still keeping every edit grounded in the trajectory that produced it. The results back it up. ReHarness beats every baseline across all three benchmarks - CL-Bench, MemoryBench, and MemoryArena - with the best overall normalized gain of 0.240, while using less than a third of the tokens that raw-trajectory baselines need. URL: #LLMAgents# #ContinualLearning#

good memory and continual learning is the next big and important inflection point in AI intelligence is important to push, but pretty saturated and you can't feel the difference between models anymore to make models more magical, they need to learn. improve. memory. supermemory

Banger paper on harness continual learning. (bookmark it) If you already are allowing your agents to rewrite their own prompts, skills, or memory files, this one is worth your time. (bookmark it) Continual learning has always tracked what changes in the weights. Modern agents accumulate experience in the harness instead, across prompts, memories, tools, skills, and routing rules. What this means is that if you update any harness component, previously reliable behavior can break with the model completely untouched. The paper names that harness-level forgetting and provides a way to measure it. Guarded harness evolution separates proposing an update from committing it. A Continual Optimizer drafts a candidate harness from post-execution feedback, and a Continual Evaluator commits only after checking current improvement, historical retention, and validity. Relative gains exceed 10% across textual reasoning, multimodal perception, and open-world interaction. Paper: Track more trending AI papers in our academy:

I wonder if Continual Learning will be solved this year. It seems to be an important stepping stone for AGI. From what I hear, Ilya Sutskever’s company SSI might be working on it. That could be super interesting if their first public release makes some progress on it.