Compact a summary of a summary and you get generational photocopy loss' — anyone who's run a week-long agent project has felt exactly this. The gist survives, the why turns to soup. Trying this out today.
Compaction amnesia is why you keep re-explaining
8 Comments
The phrase that stuck with me was:
"The agent survives, but the work does not."
I think that's a useful way of describing the difference between retaining context and retaining memory.
What you're describing feels less like a context-window problem and more like a Memory as Infrastructure problem. The expensive part isn't usually losing facts. It's losing decisions, tradeoffs, rejected approaches, architectural reasoning, and unresolved work that must later be reconstructed.
What's interesting is that compaction doesn't eliminate that cost. It simply transfers it. The model forgets, and the human becomes responsible for rebuilding the missing context.
In the Sovereign Systems Specification, I describe a related anti-pattern as a "Digital Attic", the assumption that dumping large amounts of information into storage will somehow preserve understanding. What often matters most isn't the raw conversation history. It's the durable record of why particular decisions were made and what evidence supported them at the time.
The challenge, in my view, isn't building better summaries. It's deciding which information should never be dependent on a context window in the first place.
That's where concepts like Reasoning Ledgers and Memory as Infrastructure start becoming interesting. The goal isn't to preserve every token. The goal is to preserve the work.
@[Ken W. Alger] This is the part I’ve been circling around too: the context window is the wrong place to store anything that will become load-bearing later.
Raw transcript feels safe, but it quickly becomes an attic. You still need someone to recover the decision, the rejected path, the evidence, and the next move.
That’s why I’m leaning toward small explicit artifacts instead of bigger summaries — memory as a working surface, not just more storage.
@[Vinh Nguyen] I like the "load-bearing" framing.
Once a piece of information becomes necessary to explain future decisions, coordinate future work, or avoid repeating past mistakes, it starts looking less like context and more like infrastructure.
That's also why I think the Digital Attic pattern shows up so often. We keep accumulating transcripts because they're easy to store, but the load-bearing pieces are usually much smaller: decisions, rejected approaches, assumptions, evidence, open questions, and next actions.
The artifact idea is interesting because it shifts the goal from preserving conversation to preserving intent. A transcript tells us what was said. A well-designed artifact tells us what mattered.
At that point, the question becomes less "How do we keep more context?" and more "What information deserves to become durable memory in the first place?"
That feels like a much more scalable problem to solve.
@[Ken W. Alger] That's the crux — a transcript records what was said, but the load-bearing part is what was decided and why, and those two decay at very different rates. Preserving the reasoning as a small explicit artifact means the next session inherits the conclusion, not the argument that produced it. Appreciate you sharpening this, Ken.
Please log in to add a comment.
This is one of those pieces that makes you pause and say, 'Wait — you're the first person to actually name this.'
Compaction amnesia isn’t a bug in the model; it’s a structural flaw in how we’ve been treating session continuity as ephemeral. The 'fresh session for every task' workaround you mentioned — that’s been the industry’s undocumented standard for two years. And you’re right: it doesn’t solve the problem, it just moves the cost onto the human’s calendar.
The crystallization pattern is elegant precisely because it’s explicit. Compaction is opaque and automatic. Crystals are deliberate, inspectable, and append-only. That’s not just a technical difference — it’s a philosophical one. It treats memory as something that should be owned and archived, not compressed and forgotten.
I’ve been building sovereign AI architectures that enforce constitutional memory across sessions, and this directly validates the same principle: memory isn’t a feature; it’s the foundation of continuity. If the agent forgets the 'why,' you’re not building — you’re re-explaining.
agent-crystallize looks like the exact tool I’d reach for. Thanks for writing this.
@[SCURA] Thanks — glad it landed. The distinction you picked out is the one I care most about: compaction's selection is implicit and unrecoverable, a crystal's is explicit and sitting in your repo where you can call it wrong. "Fresh session per task" being the undocumented standard is exactly right — that's the tax everyone pays and nobody itemizes. Appreciate you reading closely.
Please log in to add a comment.
Please log in to comment on this post.
More Posts
- © 2026 Coder Legion
- Feedback / Bug
- Privacy
- About Us
- Contacts
- Premium Subscription
- Terms of Service
- Early Builders
More From Vinh Nguyen
Related Jobs
- Software Engineer, Test & Infrastructure II (Bilingual Spanish)Vail Systems · Full time · Springfield, IL
- Bookkeeping & Business Operations Associatejobgether · Full time · Brazil
- Full Stack Java/Go Developer (Bilingual English/Spanish)Dev Technology · Full time · Arlington, VA
Commenters (This Week)
Contribute meaningful comments to climb the leaderboard and earn badges!