Key facts.
- SimpleQA shows even strong models frequently fail simple, verifiable questions, so an agent that re-derives facts from scratch each session repeats avoidable errors. source
- FaithfulRAG shows models override retrieved context with parametric knowledge, so even reloaded memory can be ignored unless the agent is built to honor it. source
- A cold-start agent re-derives facts each session and a bigger model still fails SimpleQA, so it repeats the rework just as wrong. (arXiv:2506.08938)
Why does cold start hurt so much?
An agent with no persistent memory treats every session as the first. It re-derives facts it established yesterday, re-asks questions a user already answered and re-makes mistakes someone already corrected, because the correction lived only in a conversation that ended. For a one-off task that is merely wasteful; for an agent meant to handle recurring work, it is corrosive, because the agent never accumulates the competence that comes from experience. The SimpleQA result sharpens the risk: models fail even simple verifiable questions, so an agent re-deriving from scratch is not just slow, it is re-rolling the dice on errors that a remembered answer would have avoided.
The fix is to give the agent durable memory: persist the corrections, the resolved facts, the user preferences and load them at session start so each run begins where the last left off. But restoring memory is only half the job. FaithfulRAG shows models will sometimes ignore the context they are handed and answer from their training instead, so a reloaded memory the agent overrides is no better than no memory at all. The agent has to be built to actually honor its persisted knowledge, treating a remembered correction as authoritative rather than as a suggestion it can overrule with a confident guess.
What does warm start require?
| Element | Cold start | Warm start |
|---|---|---|
| Prior corrections | Forgotten | Persisted and loaded |
| Resolved facts | Re-derived | Reused |
| Honoring memory | N/A | Treated as authoritative |
| Each session | Starts from zero | Builds on the last |
Persisting the right memory and getting the agent to honor it means knowing which learned facts matter for a given situation, which the Pattern Intelligence Layer makes explicit. VibeModel ties remembered corrections to the patterns they apply to, so the agent loads and respects the knowledge relevant to what it is doing, turning each session into a step forward instead of a restart from cold.
Frequently asked questions
Why not just keep the agent running?
Sessions end, processes restart and context windows fill. Durable memory loaded at start is what survives those boundaries, not an always-on process.
Is reloading context enough?
No. FaithfulRAG shows models can ignore provided context and answer from training. The agent must be built to treat persisted corrections as authoritative.
What should persist?
Corrections, resolved facts and user preferences, the things that let the next session build on the last instead of re-deriving them and risking fresh errors.