Deep memory notes
Research-backed writing on agent memory.
Practical essays on long-horizon agents, memory evals, shared behavioral guidance, and why inspectable memory becomes product infrastructure once it shapes user experience.
Why we shipped Dreaming for AI memory
The design decisions behind Background Dreaming: quiet periods, write budgets, skipped logs, revisions, and inspectable memory cleanup.
Read the essayLong-horizon agents need memory for trajectories, not just facts
Why durable memory matters when agents work across sessions, context windows, tools, corrections, and project-specific state.
Read the essayStop running evals only on prompts. Your memory changes behavior too
Prompt evals miss the behavioral surface created by shared memory, tool guides, user records, and cross-session state.
Read the essay