AI Harness Engineering · Essay Nine · derek wang (derekwang85)
Borges wrote a terrifying gift into Funes the Memorious. After Funes falls from a horse, he acquires a memory that never forgets — the outline of every leaf on every morning, the wrinkles on every face he saw in a day. He becomes a living archive so precise it borders on absurd.
But he does not become smarter. He drowns. Because he remembers every single thing, he can no longer abstract or generalize. He cannot hold the concept dog, because every dog and every sighting is a different dog. Forgetting, Borges tells us, is the quiet engine that lets us think at all; a mind that cannot forget cannot think.
That paradox is the doorway into what this essay has to say. Eight essays now have shown how to harness a single AI — steering what a model generates, restraining where that force may land. What they never solved, and the one thing that decides whether an AI system survives long-term, is simpler: can it remember what it did yesterday?
Every conversation is a fresh amnesia
Here is a scene that recurred at our company and, honestly, drove us mad. After a major incident — the "golden retriever" fraud event from Essay Eight [ORIGINAL DATA] — we held a postmortem and spelled out the lesson, clearly, out loud. The rub: the next time someone asked what actually happened in that fraud event, we had to explain it from scratch again. It exhausted the explainer, swamped the listener, and every retelling distorted the details a little more.
When the creature in that scene is human, we call it normal and lean on "team chemistry." Put the same scene in front of an AI and it is fatal: an AI's every session is, at the end of the day, the beginning of an amnesia. The moment the task ends, whatever the model implicitly learned about how we stepped in this trap and what rule we made resets along with the context window. Next time it is still the new hire on day one, ready to sink into the same hole.
Only after I understood that did I see the first eight essays for what they were: scaffolding built around a patient with amnesia. Scaffolding cannot make it remember. The real fix is long-term memory.
Knowledge is not content, it is retrieval
This is the subtle point worth picking apart. Our company did once have a knowledge system, gbrain — privately nicknamed "the dolphin." It held project knowledge, historical decisions, agent experience. It sounds complete. But the number that stuck after we migrated to derekinside:
The knowledge barely changed by a single word. Pages were 555 both before and after the migration [ORIGINAL DATA]; chunks were 2,651 both times [ORIGINAL DATA]. Nothing was added, not one page. What changed was the retrieval path.
The before/after tells the whole story:
| Metric | Before (gbrain) | After (derekinside) |
|---|---|---|
| Trigger | a human remembers to say "look it up" | AGENTS.md forces it; auto-retrieval on nearly every interaction |
| Latency | 5–10 seconds [ORIGINAL DATA] | under 1 second via HTTP API [ORIGINAL DATA] |
| Habit | queried only when remembered | injected before you think to ask |
gbrain was not short on memory or content. Its disease was reachability. The knowledge sat there, but you had to remember to go fetch it, and each fetch cost five to ten seconds. On a deadline, five seconds is a psychological hurdle and ten is a chasm, so nobody queried it. The knowledge lived on disk as a handsome exhibit of dead numbers.
derekinside did not make it smarter. It made it closer: the trigger changed from a human remembers to look to the system automatically looks, and latency dropped from ten seconds to under one. Those two moves alone woke a dead knowledge base up.
What derekinside actually is: retrieve first, then think
Ask what magic lets derekinside remember so much and the answer is almost suspiciously plain. It holds a stable internal structure — 14 wings that sort knowledge, 45 rooms that categorize it, 555 pages, 2,651 chunks [ORIGINAL DATA]. Every historical decision and every painful lesson is sliced into a small piece and placed in the right room.
The rule that makes it work does not live inside that structure. It lives in one file, AGENTS.md, and it forces every agent to run one instruction before starting:
Check derekinside first. Then think.
Not "you're welcome to check." You must check first. The trigger is automatic — when a Worker pulls a task, the knowledge base has already injected the relevant context, so you do not even have to ask. Input once, searchable forever. Put a trap, a rule, a decision in once, and on any later related task the answer reaches your hand before you have thought to ask for it.
Memory is made, not installed
Here is a real place where it slipped and climbed back, so you know it is no perfect machine. When the system swapped its embedding model, the index went stale and semantic search died outright — it could "no longer remember." Workers kept running, but every retrieval fell back to plain keywords and output quality visibly sagged.
The fix reinforces this essay's point: memory is not one-time, it is maintained. We rebuilt the index, restored retrieval, wrote the lesson into lessons/, and next time we swap a model we rebuild the index first. A system that can remember also learns to remember how to repair itself. [ORIGINAL DATA]
The counter-case: a system that remembers everything is a more expensive amnesiac
The cold water: "make the AI remember" is exactly the high-incidence zone of the Funesian trap. Funes's tragedy is remembering so much that he can no longer think. The engineering equivalent is the all-consuming knowledge base — store every lesson, archive every conversation, keep every historical version, and you get a memory palace packed with garbage. Retrieval noise climbs; what you need is unfindable and what you never asked for keeps surfacing. It looks like it remembers everything and can use none of it. That is worse than amnesia — amnesia at least does not hand you wrong answers, but a memory system stuffed with trash does.
So making an AI remember is never store more. It is remember what matters, forget what doesn't. Inside derekinside, knowledge has to be cleaned: stale decisions get flagged, outdated experience gets demoted, old conclusions give way to newer ones. Memory is alive, not a warehouse. And the testing ground for that distinction comes next — because most debugging-and-causation incidents hide in those corners of memory where we only thought we remembered.
Top comments (0)