Cue-Anchored Working Memory as a Harness Property for Coding Agents
Coding agents ship with one kind of memory: documents. Instruction files, plan artifacts, and auto-written memory directories are deliberately authored and deliberately retrieved: the agent must choose to write them and choose to read them back. Human expertise runs on a second tier that never gets written down: situationally-bound operational facts (gotchas, locations, local conventions) encoded as a side effect of the work and retrieved involuntarily when the situation cues them. We argue this second tier is the load-bearing one for long-running agents and must be a harness property, not an agent choice.
We contribute: (1) a two-tier design theory grounded in the cognitive literature on memory offloading, incidental encoding, and event-based prospective memory, each mapped to an architectural requirement; (2) 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 no surveyed academic or shipped system provides; (3) 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; (4) a repeated-compaction decay probe: ten facts held only in conversation vanish at the first summary and stay absent from 106 of 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 as the final summary carries none.
Delivery, not storage, is the product: the reliable memory channel for agents is the one the agent never has to think about.
Memory offloading, incidental encoding, and event-based prospective memory are established findings in the cognitive literature. Each one maps to an architectural requirement that document-style memory does not meet.
Every memory carries its own trigger conditions over a composable vocabulary: path, symbol, semantic, event, temporal. The harness evaluates them, so a note surfaces when its situation arrives rather than when the agent thinks to ask. No surveyed academic or shipped system provides that composition.
On a real coding task, voluntary memory use stayed at zero across 114 turns even though the answers were already in the store. Deterministic injection delivered in every seeded run with zero false alarms, and 39% of intra-session re-reads turned out to re-buy content already paid for before a compaction boundary.
Ten facts held only in conversation vanish at the first summary and stay absent from 106 of 108 compactions, at which point the deprived agent greps the harness's own session files to rebuild them. The same facts injected from a harness-owned store arrive intact through all 138 compact-resumes.
@misc{saha2026delivery,
title = {Delivery, Not Storage: Cue-Anchored Working Memory
as a Harness Property for Coding Agents},
author = {Swapnanil Saha},
year = {2026},
eprint = {2607.20972},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2607.20972}
}
Preprint, cs.AI and cs.SE, not yet peer reviewed. The runs behind every number are instrumented and written up in the two essays above, so the claims can be checked rather than taken on trust. Counter-evidence and corrections are welcome at swapnanilsaha26@gmail.com.