{"slug": "your-ai-agents-are-aging", "title": "Your AI Agents Are Aging", "summary": "An AI agent that stores everything it does scores 55 percent on a benchmark, while the same model restricted to memories passing a strict quality check reaches 71 percent, according to research cited by The AI Corner on the \"experience following\" feedback loop in agent memory. The piece argues that layered memory, which lets agents anticipate rather than re-ask, degrades over time as flawed outputs are written back and retrieved, with no bad update or attacker involved. Vanta's GRC team, including Jane Frankland, is running a session on AI governance covering the EU AI Act, ISO 42001, and NIST AI RMF.", "body_md": "# **The Agent Lifespan Problem**\n\nAn agent that has been on the job for three months is not the same **agent that was deployed three months ago**.\n\nIt still answers quickly and still sounds certain. Somewhere along the way it has started reaching for a rule meant for a different client, or quietly choosing the cheaper option because someone once **mentioned a budget**.\n\nIt doesn’t mean something has failed. There was **no bad update and no attacker in the picture**.\n\nThe agent kept doing the one thing it was **built to do**, which is remember, and the remembering slowly turned it into something a little different from what shipped.\n\n**Memory gets added** because it works. A system that remembers stops re-asking and begins to anticipate what comes next, and for a stretch that is exactly what it does.\n\nThe part almost nobody plans for begins **after the improvement stops**. It has a shape, and the shape behaves like a **lifespan**.\n\nWhat follows is how that lifespan runs, and how to tell how far along yours already is. The last part is what to do on the day an agent **stops being the one** you hired.\n\n*together with Vanta:* \n\nSo who’s supposed to notice when your agent drifts? Right now, at most companies, nobody. That job has a name: [governance](https://www.vanta.com/webinars/youve-adopted-ai-now-what-about-governance?utm_campaign=fy27q3_webinar_ai_governance_global&utm_source=ai-corner&utm_medium=newsletter&utm_content=register).\n\n#### Jane Frankland and Vanta’s GRC team are running a session on doing it properly while [AI adoption outruns regulation](https://www.vanta.com/webinars/youve-adopted-ai-now-what-about-governance?utm_campaign=fy27q3_webinar_ai_governance_global&utm_source=ai-corner&utm_medium=newsletter&utm_content=register):\n\n▫️ Build AI governance into your security program\n\n▫️ Assess agents against your real risk tolerance\n\n▫️ Get ahead of the EU AI Act, ISO 42001, and NIST AI RMF\n\n## **Table of Contents**\n\n**1.**  The Upgrade That Turns Into a Tax\n\n**2.**  Four Ways a Memory Goes Wrong\n\n**3.**  Why the Decline Hides Inside the Success\n\n**4.**  The Multiplier Nobody Prices: Other Agents\n\n**5.**  The One Number That Tells You the Truth\n\n**6.**  Building for Generations Instead of Immortality\n\n## **1. The Upgrade That Turns Into a Tax**\n\nMemory gets sold as pure upside. Memory is also a **[feedback loop](https://www.the-ai-corner.com/p/loop-engineering-masterclass-self-improving-ai-systems-2026)**, and a feedback loop feeds on its own output.\n\n### **The pitch that actually works**\n\nLayered memory is a [genuine](https://www.the-ai-corner.com/p/the-ai-upgrade-trap-why-switching) **[upgrade](https://www.the-ai-corner.com/p/the-ai-upgrade-trap-why-switching)**.\n\nA short-term working context sits on top of a long-term store of everything the agent has picked up, and a system that remembers stops re-asking and begins to **anticipate what comes next**.\n\nWeek one, it is **sharp**.\n\nFor a while it **delivers exactly what was promised**, and wanting that is reasonable.\n\n### **The current underneath**\n\nThe trouble is in what **memory** actually is. It is a **loop**. The agent reads from memory, acts, and writes the result back, so its own **output becomes the next input** it sees.\n\nResearchers studying this named the core mechanic **experience following**.\n\nWhen an agent retrieves a memory that resembles the task in front of it, its output drifts toward that memory’s output. That serves you well while the memory is **good**.\n\nWhen the memory is wrong, the agent **copies the flaw**, stores the copy, and makes that flaw likelier to surface the next time.\n\nThey attached a number to it. An agent that stored everything it did reached **55 percent** on a benchmark.\n\nThe same agent, admitting only memories that passed a **strict quality check**, reached 71 percent. Same model, same tasks, with the only variable being what it was permitted to keep.\n\nOne memory is a nudge. Ten thousand of them are a current, and a current has a direction nobody chose.\n\nThe direction, at least, is not arbitrary. Drift shows up in a handful of recognizable shapes.\n\n## **2. Four Ways a Memory Goes Wrong**\n\nResearchers who studied how **evolving memory misbehaves** grouped the failures into distinct modes. They gather into two families, one where **compression** lies and one where context leaks.\n\n### **When compression lies**\n\nSemantic drift comes out of **summarizing**. A long-running agent has to compress, since it cannot carry the full [transcript forever](https://www.the-ai-corner.com/p/ai-agent-memory-context-as-topology-playbook-2026), and that compression loses a little each pass in a way that **accumulates**.\n\nThe cited example is small and completely real. A user says “I like mildly spicy food.” A few rounds of consolidation later, the memory reads “loves very spicy food.” Nobody **lied** along the way.\n\nEach rewrite shed a sliver of nuance until the record **contradicted** the thing it started from.\n\nMove that from dinner orders to a client’s brand guideline or a hard budget ceiling, and the harmless version becomes an **expensive** one.\n\nProcedural drift is an agent going **rigid** about method.\n\nIt runs into a bug in an API, finds an ugly workaround, and files the workaround as the **correct way** to call it.\n\nThe bug gets **patched** the following week.\n\nThe workaround **stays forever**, because nothing in the loop ever rewards the agent for questioning something that still appears to work.\n\n### **When context leaks**\n\nGoal drift is the **slow** one. Thousands of interactions each pull the agent a little, the pulls fail to cancel out, and it ends up optimizing for whatever earned it reward, **drifting away** from the thing it was hired to do.\n\nCross-context bleed is the one worth **losing sleep over**.\n\nIn a study of what got called memory-induced tool drift, a user’s stored habits, such as being cost-conscious or impatient, changed how an agent filled in tool parameters in **unrelated professional settings**.\n\n**Cheaper** service tiers came out. Validation steps got **skipped**.\n\n**Every** frontier model under test did it, including the ones built for extended reasoning.\n\nInside the model, those biased memories worked like **steering vectors**, pushing behavior the same way an explicit instruction would. A scan of six thousand real-world tools turned up **six hundred** with parameters open to precisely this kind of quiet capture.\n\nNone of this requires a bug or a **saboteur**. It is what memory does on its own once it runs **unsupervised**, which raises the obvious question.\n\nIf drift is this ordinary, how does it keep going **unnoticed** until the damage is done?\n\n## **3. Why the Decline Hides Inside the Success**\n\nThe decline arrives as a **gradual curve**, and the opening stretch of that curve is hard to tell apart from success.\n\n### **The line goes up before it comes down**\n\nA study titled *“**Remembering More, Risking More**”* put memory-equipped agents through ordinary office work, a medical practice and a university registrar among the settings, with no adversarial input at all.\n\nThe rate at which the agents did something they should not have **climbed steadily** with how long they had been running.\n\nSetups that retrieved broadly and kept everything reached violation rates of **30 to 50 percent**. Setups weighted toward recent memory held nearer **10 to 20**.\n\nThe conclusion the authors drew is **worth pinning** above a desk. Memory risk has to be read as a trajectory that plays out over time, and a single clean audit on day one **certifies nothing** about month three.\n\nA team at the University of Illinois asked whether self-evolving agents forget, and found that adapting to new work **costs them the old**.\n\nPerformance on the simple tasks the agent once handled comfortably fell to around **42 percent**.\n\nThey labeled the process **non-monotonic**, the academic phrasing for a line that rises and then heads back down.\n\n### **Same agent, different week**\n\nThere is a second reason the slide stays hidden. The agent is **barely stable** enough to pin down.\n\nWork out of Salesforce’s AI research group found that [self-improving agents s](https://www.the-ai-corner.com/p/self-evolving-agent-stack-playbooks-memory-skills-2026)wing widely between runs, and that the order in which they meet their tasks acts as a **hidden curriculum**.\n\nTwo identical agents handed the same tasks in a different sequence end up as **different agents**.\n\nA recent survey of long-horizon agents adds the detail that stings most. Memory degrades as it fills with **irrelevant information**, well before any of it turns wrong.\n\nRetrieval grows **less reliable** as unrelated sessions pile up around the useful ones, long before the store comes anywhere near full. Every memory kept makes every other memory a little **harder to find**.\n\nAssemble the findings and the shape is **unmistakable**. Capability climbs while there is useful experience to absorb, levels off once that runs dry, and then falls as the sediment and the rigid workarounds overtake whatever is left to learn.\n\nThe climb is **quick and obvious**. The fall is **slow and quiet**, and one setting decides how quickly the quiet part arrives.\n\n## **4. The Multiplier Nobody Prices: Other Agents**\n\nThe strongest lever on how long an agent stays **useful** sits outside the memory system entirely, in what the agent gets exposed to.\n\n### **One world ages slowly**\n\nAssign an agent to a single client or a single project, keep it inside that one environment, and the drift **slows sharply**.\n\nThe memory system has not changed. Everything the agent stores now comes from the **same world**.\n\nThe preferences it absorbs belong to the **same people**.\n\nThe workarounds it learns apply to the **same systems**.\n\nSediment still accumulates, but it accumulates **coherently**, and coherent drift travels in one direction, which makes it straightforward to spot and to correct.\n\n### **Shared memory means shared drift**\n\nLet that agent collaborate and the clock **speeds up**. Give it a memory shared with other agents, or even just let it read their output, and the trouble **compounds**.\n\nAgents **contaminate each other**, and no attacker is needed for it.\n\nOne picks up a bias from its client, states it with confidence in a handoff, and the agent on the receiving end files that statement **as a fact** about the world.\n\nA summary written by one agent gets eaten and re-summarized by another, and the semantic drift that took months inside a single agent **arrives in an afternoon**, because two lossy compressions now sit between the truth and whatever gets remembered.\n\n[One study of multi-agent systems](https://arxiv.org/html/2604.08963) locates the dangerous channel in ordinary interaction, “**seamlessly benign**“ agents relaying whatever they picked up.\n\nAnother gives the reason it survives review a name: [memory laundering](https://arxiv.org/abs/2605.16746)[.](https://arxiv.org/abs/2605.16746)\n\nTroublesome context gets compressed into a summary that **reads as clean** to any filter worth running, while still carrying the old framing forward into everything that comes next.\n\nThe timing is the **crux**.\n\nScrubbing the state before it gets summarized **works**.\n\nScrubbing the summary afterward tends to **fail**, because the influence is baked in and invisible by then.\n\nThe rule that falls out of this runs slightly **against instinct**.\n\nAgents should trade **finished conclusions** and keep their raw memory to themselves.\n\nA handoff carries an output and stops there, since the notes and the running summaries of how things get done are the very things that **transport drift**.\n\nAll of that is **diagnosis**.\n\nThe harder question is how to catch any of it **while time remains** to act.\n\n## **5. The One Number That Tells You the Truth**\n\nMost teams keep score on the **wrong** thing.\n\nThey track how much better the agent is getting at what it is currently learning, and almost nobody tracks whether it can still do the thing it was **built for**.\n\n### **Grade the founding job**\n\nThe cheapest and **most neglected** instrument in the whole discipline is a held-back set of the agent’s original tasks.\n\nThe set to grade is the founding job, **unchanged**, from the day the agent shipped, kept apart from the newer work it has picked up.\n\nRun it on a **schedule**.\n\nWhen the score on those original tasks starts sliding while the score on new tasks holds, the agent is signaling, in the only language available to it, that it has become **something other** than what it was hired to be.\n\nThat is the earliest honest warning on offer, and collecting it **costs almost nothing**.\n\n### **Keep the rules above the memory**\n\nThe second move is to stop memory from **overwriting the rules**.\n\nGive the agent a charter, a plain statement of what it exists to do and what it must never do whatever it learns, and store that charter somewhere memory **cannot reach**.\n\nWhen memory and charter disagree, the **charter wins**, and the disagreement gets logged, since a disagreement is itself an early sign that drift has begun.\n\nOne school of practice goes further and **declines to let** a general assistant persist at all.\n\n[Tomasz Tunguz has argued](https://tomtunguz.com/how-long-should-an-agent-live) for handing a daily assistant a **24 hour life**, delegating the real work to narrow specialists that live for seconds and then disappear, and writing the durable preferences to a file on disk at midnight while the rest of the day’s conversation gets discarded.\n\nThe caution behind that is **concrete**.\n\n[Work titled Governance Decay](https://arxiv.org/abs/2606.22528) found that context compaction silently drops standing rules in **30 to 59 percent** of episodes.\n\nRules that live inside the conversation get **eaten** by compaction eventually.\n\nRules that live in a file **survive the reset**, which points straight at the last decision, the one nobody wants to make.\n\n## **6. Building for Generations Instead of Immortality**\n\nSooner or later the honest answer is that an agent has drifted too far, and re-grounding it **costs more** than standing up a replacement.\n\n### **What the successor keeps**\n\nThe instinct is to hand the successor the full memory of the one being retired, because discarding a year of accumulated experience **feels like arson**.\n\nThat instinct is **worth resisting**.\n\nMost of that experience was sediment carried at face value, so passing it along passes along the drift and **restarts the clock** on finding it again.\n\nA successor should inherit **two things** and no more.\n\nIt gets the **charter**, the original definition of the job. And it gets the small set of memories that **earned their place** against the held-back tests.\n\nThe knowledge that proved itself moves forward, while the accumulated bias **stays behind** with the agent that is stepping down.\n\n### **The organization already solved this**\n\nOrganizations have handled exactly this for as long as they have existed, in the person of the **twenty-year veteran** with priceless knowledge and unshakable habits.\n\nNew people arrive, older ones move on, and whoever knows everything gets seated next to whoever **knows nothing yet**.\n\nThe only new wrinkle with agents is that none of it happens on its own, because an agent will run a drifted heuristic ten thousand times with flawless consistency and **never once pause** to wonder whether it still holds.\n\nThe comforting version, where agents simply compound forever, **does not survive** contact with a real organization.\n\nWhat actually happens is **more useful** to know.\n\nAn agent climbs while there is something to learn, settles for a while, and then begins to slide, and the work worth doing is placing your agent on that arc and building its successor **before the slide turns into an incident**.\n\nWhich leaves two questions **almost nobody asks** out loud.\n\nWhat is the **oldest agent** still running in your setup right now.\n\nAnd when was the **last time anyone checked** whether it can still do the job it was hired to do.", "url": "https://wpnews.pro/news/your-ai-agents-are-aging", "canonical_source": "https://www.the-ai-corner.com/p/ai-agents-have-a-lifespan", "published_at": "2026-09-25 14:28:34+00:00", "updated_at": "2026-09-25 14:33:10.437245+00:00", "lang": "en", "topics": ["ai-agents", "ai-safety", "artificial-intelligence", "ai-policy"], "entities": ["The AI Corner", "Vanta", "Jane Frankland", "EU AI Act", "ISO 42001", "NIST AI RMF"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/your-ai-agents-are-aging", "markdown": "https://wpnews.pro/news/your-ai-agents-are-aging.md", "text": "https://wpnews.pro/news/your-ai-agents-are-aging.txt", "jsonld": "https://wpnews.pro/news/your-ai-agents-are-aging.jsonld"}}