When an AI picks up where earlier steps left off, does it reread everything or just a curated slice?

This explores what happens mechanically when a later step, turn, or task in an AI system picks up what earlier ones left behind: does it re-read everything, or work from something narrower? Across the corpus the answer is consistent. Later work almost never acts on the full inherited history. It acts on a curated slice, and the interesting design question is who or what does the curating.

The first approach is to hide most of it. LLM Programs put an ordinary algorithm in charge of state and control flow, and each model call sees only the context relevant to its own step Can algorithms control LLM reasoning better than LLMs alone?. The Thread Inference Model does something similar inside a single model. Reasoning is organized as a tree of subtasks, and when one finishes, fixed rules prune its cache entries so the parent carries forward only what it needs. This keeps reasoning accurate past the context limit, even while manipulating 90% of the cache Can recursive subtask trees overcome context window limits?. Atom of Thoughts goes furthest. Each reasoning state is rebuilt to depend only on the current, simplified problem, not on the path that produced it Can reasoning systems forget history without losing coherence?.

The second approach is to pass state by reference. ReWOO and Chain-of-Abstraction have later steps act on earlier results through placeholders. The plan is written first and the tool outputs are slotted in as they arrive. That avoids replaying every observation into every prompt, which otherwise makes prompts grow quadratically Can reasoning and tool execution be truly decoupled?. Chain-of-thought is the opposite extreme. A feedforward transformer has no native way to carry evolving state forward, so it writes that state out as tokens so the next step can read it. That works, but it is a costly patch Why do transformers need explicit chain-of-thought reasoning?. A related finding recasts the long-context problem as one of compute. What limits performance is the work needed to turn evicted context into internal state, and more consolidation passes help Is long-context bottleneck really about memory or compute?.

The third approach is to gate what gets stored, then be selective about how it is read back. Long workflows degrade when agents replay transcripts or retrieve memory with no control over writes. The fix is a small, schema-governed committed state that separates temporarily recalling an artifact from permanently writing it to memory, which stops errors and drifting constraints from piling up Can agents fail from weak memory control rather than missing knowledge?. On the read side, PRAXIS indexes procedures by the current environment state and the action taken there. A later web task pulls the click-level step that matches where it is, not a vague workflow summary Does state-indexed memory outperform high-level workflow memory for web agents?. MRAgent skips fixed retrieval and rebuilds memory on demand. It walks a memory graph while reasoning and prunes paths as evidence accumulates, gaining up to 23% on reasoning tasks Can agents reconstruct memory on demand instead of retrieving it?.

Inherited state also includes rules, and this is where the safety angle appears. A check that looks only at the current action cannot state a constraint that depends on history, so it misses actions that are each fine alone but together break a system-level rule. Catching those takes a stateful monitor Can stateless checks ever catch sequence-level constraint violations?. One persistent agent logged 889 governance events over 96 days because its safeguards lived in the memory layer it actually consulted while deciding Can governance rules embedded in runtime memory actually protect autonomous agents?. Inherited state only shapes later behavior if it sits where the next step looks.