Memento Principle¶
matt-pocock's name for an LLM workflow heuristic: treat the agent like Leonard from Memento — it forgets by design, so reset the session instead of compacting it.
"Devs love compacting for some reason, but I hate it. I much prefer my AI to behave like the guy from Memento because this state is always the same. Always the same every time you do it. You clear and you go back to the beginning." — Pocock, 10:36–10:50
The argument¶
Compaction (/compact in claude-code) squeezes a long conversation into a written summary and continues the session. Pocock's objection: the resulting "post-compact" state is non-deterministic and already polluted — sediment from prior phases lingers in the summary and biases future tool calls. By contrast, /clear returns to a known, reproducible system-prompt-only state — the same starting point every time — which is easier to reason about and keeps you in the smart zone.
Workflow implication: plan in long-lived documents (PRD, issue files — see prd-as-destination) that can be re-read into a fresh session, rather than relying on conversational memory to persist through compaction.
Tension with existing wiki¶
This directly contests the framing in context-compaction from the Zoneraich ingest (jared-zoneraich-claude-code-works-2026), which describes H2A auto-compaction at ~92% as a load-bearing mechanism inside Claude Code's main loop. Both views can coexist: H2A keeps long agent-driven loops alive; Pocock's principle applies to human-driven coding sessions where clearing is cheap because the plan is already externalised.