Abstract
Coding agents come equipped with one type of memory: documents. Instruction files, plan artifacts, and auto-generated memory directories are deliberately authored and retrieved; the agent must choose to write and read them. Human expertise operates on a second tier that never gets recorded: situationally-bound operational facts (gotchas, locations, local conventions) encoded as a byproduct of work and retrieved involuntarily when cued by the situation. We argue that this second tier is the load-bearing one for long-running agents and must be a harness property, not an agent choice.
We contribute:
- A two-tier design theory grounded in cognitive literature on memory offloading, incidental encoding, and event-based prospective memory, each mapped to architectural requirements;
- A cue-anchored memory model where memories carry first-class trigger conditions over a composable vocabulary (path, symbol, semantic, event, temporal), evaluated deterministically by the harness, a composition not provided by any surveyed academic or shipped system;
- A controlled evaluation on a real coding task showing that voluntary memory use is near zero even with a pre-seeded store (0 memory operations in 114 turns), that deterministic injection delivered in every seeded run with zero false alarms, and that 39% of intra-session re-reads re-buy content paid for before a compaction boundary;
- A repeated-compaction decay probe: ten facts held only in conversation vanish at the first summary and stay absent from 108 compactions, and the deprived agent greps the harness's own session files to rebuild them, while the same facts injected from a harness-owned store arrive intact through all 138 compact-resumes with the final summary carrying none.
Delivery, not storage, is the product: the reliable memory channel for agents is the one the agent never has to think about.
Blogger's Review: This study reveals the potential advantages of coding agents in working memory, emphasizing the significance of situationally-triggered memories in long-term tasks. By introducing a two-tier design theory and cue-anchored model, the research provides new insights for achieving more efficient agent systems, which is worth attention.