Memory Is Not a Warehouse but an Alarm: When AI Learns to "Remind When It Matters, Stay Silent When It Doesn't"
A Feynman-style paper interpretation, originally posted in Chinese on zhichai.net.
> *"Memory is the monument of the soul, but it only has meaning when it is read."* > *— adapted from Borges*
---
A Familiar Scenario
Imagine playing a complex puzzle game with 100 rooms. In room 3 you find a note: "The treasure room passcode is the number of vases in the room." Later, room 17 has one vase, room 42 has two, room 89 has none.
Five hours later, standing at the treasure room door, you enter... "3." Wrong. You scramble through your notes — but you never wrote it down, because "how could I forget something so simple?"
This is Behavioral State Decay — the paper's core concept. The information still exists somewhere, but at the exact moment it is needed, it fails to surface. AI agents in long-horizon tasks face exactly the same problem.
---
The Agent's "Goldfish Memory" Dilemma
Picture an agent asked to: log into a server, find last month's error logs, analyze the cause of a crash, and write a repair report.
1. It logs into the server ✅ 2. It finds the log directory ✅ 3. It discovers the logs are compressed and decompresses them ✅ 4. It analyzes and finds the crash may relate to a config file 5. The config file needs sudo; sudo needs two-factor authentication...
By this point, does the agent still remember the credentials from step 1, the exact log path from step 3, the suspicious timestamp from step 5? As the trajectory grows, decision-relevant information sinks deeper into the context stack — or gets pushed out of the context window entirely.
---
Core Insight: Memory Should Be Active, Not Passive
The conventional fix — a bigger context window or an external memory store — is like wallpapering your house with sticky notes. But information overload is itself the problem: an agent at step 100 that must juggle steps 1–99 has its attention diluted past the point of making good decisions.
The paper's insight:
> Memory should not be a passive warehouse ("fetch it when you remember you need it") but an active alarm bell ("it rings exactly when you should remember").
---
Architecture: A Memory Agent That Reads the Room
Rather than modifying the Action Agent, researchers give it a parallel Memory Agent that behaves like a skilled assistant:
1. Observation
It continuously "watches" the Action Agent's work.2. Memory Bank Update
It extracts structured information from recent trajectories into a memory bank:- What are the task requirements?
- What are the environment's characteristics?
- Did previous attempts succeed?
- What problems were diagnosed?
- Which subgoals remain unfinished?
- Terminal-Bench 2.0: +8.3 percentage points over the baseline without a Memory Agent
- τ²-Bench: +6.8 percentage points
- ❌ Passive memory bank exposure (dump everything into context): worse — information overload dilutes attention.
- ❌ Always-on injection (reminders every few steps): even worse — like an alarm that nags you every five minutes; you'd smash it.
- ❌ Advisor-only mode (suggestions without concrete memory content): limited gains — the agent needs "the config file is in /etc/nginx/", not "you might check the config".
- ❌ Generic retrieval: underperforms — the agent doesn't know what to retrieve, which is the very definition of behavioral state decay.
- ✅ Selective intervention: best. The Memory Agent judges when to remind and what to remind about.
- 🎹 The conductor's assistant: holds the score, coughs softly only when the conductor seems about to miss a page turn.
- 🍳 The sous-chef: warns "this dish is low-salt" just as the chef reaches for salt; says nothing when it's the sugar jar.
- 🧭 The mountaineering support team: the radio stays silent during normal climbing, then crackles: "you are approaching the ice crevasse area."
- Correct reminder at the right time → high reward
- Unnecessary interruption → negative reward
- Missing a needed reminder → negative reward
- Wu, Y., Zhang, L., Zhou, Y., et al. (2026). *Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents*. arXiv:2607.08716.
- Wang, X., et al. (2024). *Terminal-Bench: A Benchmark for Terminal-Based Agents*.
- Team, Q. (2025). *Qwen3.5 Technical Report*.
- Borges, J. L. (1941). *The Garden of Forking Paths*.
- Feynman, R. P. (1964). *The Feynman Lectures on Physics*.
3. Selective Intervention: Speak or Stay Silent?
The Memory Agent decides: *does the Action Agent need a piece of information it may have forgotten, right now?*
If yes → it injects a memory-grounded reminder. If no → it stays silent.
Like a good assistant: she warns you about a trap clause just before you sign a contract — but doesn't interrupt every five minutes to offer coffee while you're coding.
---
Results: Selective Intervention Beats Everything
Crucially, the Memory Agent helps both weaker and stronger Action Agents — the benefit of memory intervention is universal, not dependent on the underlying agent's capability.
---
Ablations: Why "Selective" Is the Key
---
Analogies
A good assistant: knows when to speak, knows what to say, knows when to stay silent.
---
Training: Teaching Agents to Read the Room
The researchers trained a Qwen3.5-27B model with SFT + GRPO to learn the "when to intervene" policy, using a dataset called SETA (Selective Event Triggering for Agents).
"When to remind" isn't a question with a single correct answer — it's a policy problem: too early disturbs, too late misses the moment. GRPO's reward function:
Preliminary results: improved validation reward and partial transfer to Terminal-Bench. In other words, selective memory intervention is a learnable skill.
---
Deeper Meaning: From "Storage" to "Curation"
Traditional AI memory systems treat memory as a storage problem: how to store more, retrieve faster. This paper reframes memory as a curation problem.
Think of a museum curator: she doesn't cram every artwork into a warehouse. She decides what to collect, when to exhibit what, how to present it, and when a piece should "rest." The Memory Agent is that curator — presenting the right information, in the right way, at the right time.
Curation, not storage.
---
Looking Forward: From "Tool" to "Partner"
Future AI assistants may no longer be "ask a question, get an answer," but rather: quietly watching you work, gently reminding you when you're about to err, supplying key information when you forget, and not interrupting while you focus. This "silent companionship" may be closer to what humans truly want from an intelligent assistant than any flashy feature.
---
Conclusion: The Art of Memory
Every agent interaction forks a path through time. The Memory Agent's purpose is not to remember every path — that's impossible — but to let the agent remember why it set out on each one.
Behavioral state decay is not memory loss; it is memory asleep. The Memory Agent's job is to wake those sleeping memories — at the right time, in the right way.
> The value of memory lies not in possession, but in recall.
---