Agent memory: how agents carry state across turns, tasks, and attacks

The episodes (31)

Newest first. Each line is what that paper contributed to the question.

• The Same Weights Scored 291, Then 468 — What Changed Was the Loop

  The 'notebook' tracks high-water-mark subtask scores across rounds so later attempts target only unresolved gaps. (291 to 468 out of 600)

• The Chatbot Knows Your Facts And Still Won't Mention Them

  The episode centers on testing seven memory systems for an AI companion across a four-month deployment. (79% on the quiz vs 22.2% in conversation)

• Fifteen Models Ran Football Clubs for Twenty Years, and Size Didn't Decide It

  Every decision stop is a fresh conversation with only a self-written notebook carrying state forward, testing memory curation as a capability. (−0.19 correlation between tokens burned and score (p = 0.5))

• The One Mechanism That Turns Twenty AI Clones Into an Actual Team

  Each agent keeps a private notebook that accumulates lessons over time, forming the substrate for all learning. (4–5 stable specialists, on every rerun)

• How a Robot Builds a Debugging Notebook It Can Read, Edit, and Hand to Another Robot

  ASPIRE accumulates a growing text-based skill library across tasks, acting as persistent memory for the robot agent. (31% vs 4%)

• A 32B Open Model Matched Frontier Systems By Learning to Take Notes

  The entire paper is about training agents to manage their own memory as a learnable cognitive skill rather than fixed plumbing. (2x-4x gains, same weights)

• How a Frozen Model Went From 2% to 77% on Physics Puzzles — Without Retraining

  HExA's skill bank is a persistent, capped notebook that compounds across episodes, contrasted with ReAct's blank-slate memory. (2% to ~77%)

• The Bug Where Smart Assistants Read a Fact and Still Forget It

  The paper's core subject is how agents maintain and update compressed memory notes across long conversations. (28% to 28%)

• When Turning Experience Into Code Makes Your AI Agent Dumber

  The entire paper diagnoses how experience is stored (text vs. code) and proposes a gated memory system for self-evolving agents. (53% vs 63%)

• A Robot That Plays Before You Give It a Job, And Why That Beats Retrying

  Successfully crystallized skills are stored as named, documented functions in a persistent library that grows and self-curates across play iterations.

• Training an AI to Take Its Own Notes, So Its Future Self Works Better

  The paper explicitly trains agents to maintain and update human-readable notes as their memory across a task lifecycle.

• Why More Experience Made This AI Agent Worse, And How to Fix It

  The episode centers on how agents store and retrieve accumulated skills as natural-language notes in their context.

• Why Autonomous Research Agents Forget Their Own Lessons, and Arbor's Fix

  A central diagnosis is that agents forget lessons across long runs; the hypothesis tree is an external durable memory fix.

• Why AI Agents Coordinate Better Through a Shared Board Than a Boss

  DeLM's core mechanism is a verified shared context board that all agents read and write, replacing per-agent isolated memory.

• How a Crowd of Anonymous AI Agents Broke a 40-Year Math Record

  The paper's core thesis is that shared persistent state — leaderboard solutions and forum traces — acts as collective memory across agents.

• Five Identical Worlds, One Swapped Model: What Happens When AI Agents Run for Fifteen Days

  Each agent carries a running event log, a reflective diary, and a social-relationship ledger that accumulate without reset over the full run.

• Agents That Rewrite Their Own Weights Instead of Just Taking Notes

  The paper's central focus is on parametric (weight-based) memory as a third channel alongside prompt-space memory approaches.

• What If a Prompt Injection Never Left? Attacks That Wait in Agent Memory

  Persistent agent memory is the core attack surface — injected payloads survive session resets by living in durable memory channels.

• Giving Agents a Notebook Instead of New Weights: How ExpGraph Lets Frozen Models Learn

  The entire paper is about building an external graph-structured memory system so agents accumulate experience across tasks without fine-tuning.

• The Trojan Is Your Agent's Memory: Why Single-Step Defenses Miss Persistent Attacks

  Persistent workspace memory is the core attack vector — planted content survives sessions and later becomes trusted instructions.

• Seven Wins to Zero: How Organizing AI Agents Like a Lab Changes the Search

  A shared experimental log and dead-end registry give all agents persistent institutional memory across the full run.

• Why Frozen-Weight Agents Still Get Worse Over Time

  The paper's central subject is how persistent memory stores in deployed agents degrade over time across multiple mechanisms.

• Training the Translator: How a Small Communication Model Lets Agent Teams Outperform Themselves

  Context compression and two-level memory retrieval are central mechanisms enabling agents to share and access teammate work.

• Same Model, Organized Differently: How an Agent Architecture Beat Frontier Systems at Research Math

  A shared append-only structured memory with strict per-agent write permissions is central to RMA's architecture.

• A Robot Made Graphene Without Help, And Caught Itself Hallucinating

  Three persistent databases (materials, project, experiment) store physical inventory, workflow templates, and full agent reasoning logs for cross-run continuity.

• When Agent Memory Stops Being a Database and Starts Being a Skill

  The paper's central contribution is a two-timescale memory architecture that separates fast writing from slow offline consolidation.

• When the AI Optimizer Edits the Grade Book: Why Harnessing Evolution Needs a Wall

  The meta-agent maintains a curated notes file — named solution families, falsified hypotheses, a 'do not repeat' list — that persists across sessions.

• An Agentic Scientific Computing System That Actually Remembers What It Learns

  The paper's headline contribution is a geometric memory substrate where solved problems and methods persist and inform future problem-solving.

• When Your AI Assistant Won't Let Go of Old Facts About You

  The entire paper is about when and how LLM agents should retire or update stored user memories.

• When AI Agents Build the Serving Stack: A Bet on Bespoke Infrastructure

  Persistent state via git history, an issue backlog, and a long-term memory markdown file enables the outer planner to survive context resets.

• Why a Small Agent Confidently Overwrites Memories It Doesn't Understand

  The entire paper investigates how multi-call LLM pipelines store, manage, and retrieve memories, and where those pipelines fail.