Working Memory
Ideas I haven’t finished thinking about.
Field notes from building agentic systems and half-formed arguments — loosely held, heavily interlinked, and updated as they grow.
- 03 Oct'26
Context windows are a budget, not a bucket
Every token spent on history is a token not spent on reasoning. Treat the window like a budget line, not storage.
- 01 Oct'26
Memory is a product decision
What an agent remembers, forgets, and surfaces is a product surface — not a side effect of the vector store you picked.
- 30 Sep'26
Write the eval before the feature
An eval is where a team writes down what "good" means. Do it first and the feature designs itself.
- 28 Sep'26
Latency is a UX problem wearing an infra costume
Users don't experience p95 — they experience waiting without feedback. Most "latency" wins are UX, not infra.
- 25 Sep'26
The harness is the moat
Models commoditize. The scaffolding around them — tools, memory, evals, recovery — is where the defensible product lives.
- 20 Sep'26
Tool calls are I/O, so treat them like I/O
Latency, retries, timeouts, idempotency — the discipline we learned for network calls applies directly to agent tools.