本週人人都在構建智能體記憶,卻未就係統該忘記甚麼達成共識
Everyone Is Building Agent Memory This Week. Nobody Agrees on What It Should Forget.
作者指出,智能體記憶正成為熱門議題,但系統如何處理過時或誤導記憶,仍是尚未解決的設計難題。DEV Community 過去 7 天的趨勢頁列出多組相關集羣,其中 SRE 智能體事後記憶有 83 篇,具歷史記憶的值班智能體有 72 篇。文章主張為記憶保留來源、時間與適用條件,並把它視為可供當前事故質疑的證據,而非禁止重試的清單。
Open this site's trending page and play a drinking game: take a sip every time a trend cluster is about AI agent memory. You will be hospitalised. This week the page carries cluster after cluster on the same idea — hindsight memory for SRE agents (83 posts in the last 7 days), on-call agents with historical memory (72), agent memory and provenance in incident response (69), hindsight agents in incident response (60), human-gated agent workflows (58). Different titles, one obsession:
Agents should remember what already failed, so they stop suggesting it at 3 a.m.
It's a good obsession. A stateless assistant that watches you restart a connection pool during an outage will suggest restarting the connection pool during the next outage too, forever, with fresh enthusiasm. Memory is the obvious fix. Remember the postmortem, remember the rollback that made it worse, remember the fix that worked. Feed it back in.
But read the trend clusters' own question lists — the site generates them from the posts — and a second question keeps surfacing underneath the first: what stops the agent from acting on a stale or misleading memory? That one has no cluster of its own yet. It's the harder half of the problem, and I think it's where this gets decided.
Memory is just more input
I write about agent security here, so I have a reflex about this: everything an agent reads is a surface somebody can write on. A memory store is a file, or a database, or a pile of postmortems, that the agent will treat as ground truth about the past. Postmortems get edited. Runbooks rot. A "failed fix" recorded during one incident might have failed because the incident was misdiagnosed — the fix was fine, the memory is a lie the system tells itself with total confidence.
The teams in these trend posts are not naive about it; the phrases "contradiction guards" and "provenance" show up constantly. Provenance is the good instinct: a memory should carry where it came from and when, the way a receipt carries a date. A fix that failed in March against a different database version is not evidence about tonight. Without the date and the source, though, memory collapses into folklore — we don't do that here, nobody remembers why — except the folklore is now machine-readable and delivered at machine speed.
The forgetting problem is the design problem
Humans forget, and we mostly treat that as a defect. For agents it might be a feature we're about to engineer away by accident. Consider what an agent with perfect negative memory does in month six: it has a long list of things that failed once, under conditions nobody recorded, and it refuses to try any of them. Some of those things failed because they were wrong. Some failed because the timing was wrong, the deploy was half-finished, the person running them was tired. An agent that can't re-try a once-failed fix under new conditions isn't experienced. It's superstitious.
The interesting designs in this week's posts are the ones that treat memory as evidence with a weight, not a ban list: this failed twice, under these conditions, recorded by this operator. Retrieve it, show it, let the current incident argue with it. That's closer to how a good senior engineer uses experience — we tried that in March and it bit us, here's the postmortem, check whether tonight matches — and it's a much harder system to build than a vector store with a grudge.
What I'd watch for
Three questions I'd ask any agent-memory system, including the ones flooding this site:
- Can a memory be wrong in a way the system can notice? If two memories contradict each other, does anything flag it, or does the agent just average them into confidence?
- Does the memory know its own date and source? "Failed fix" without provenance is gossip.
- Can the agent ever overrule its memory, and does it leave a note when it does? The override, recorded, is how the memory gets corrected. Without it, month one's mistakes are month twelve's policy.
The memory wave is real and it's pointed at a real problem — I've watched my own agent experiments repeat a bad suggestion because nothing in the setup remembered the last failure. But the systems that win this won't be the ones that remember the most. They'll be the ones whose memories can be argued with.
If you're building one of these: what does your agent do when its memory is wrong? I'm genuinely asking, because that answer is the product.
P.S. Everything I write stays free — that part doesn't change. But if this one was worth something to you, you can buy me a coffee ☕ (one-time or monthly, both work). It covers the API credits and hosting behind these experiments — and honestly, it's also the nudge that gets the next one started on the evenings I'd otherwise skip it. — Salam
來源:DEV Community · dev.to
