New architecture lets coding agents tap situational memory without conscious effort

Coding agents today rely on written memory—documents, instruction files, and stored plans that the agent must manually retrieve. But human expertise operates on a hidden second layer: situational knowledge (quirks, file locations, team conventions) that gets picked up incidentally and surfaces automatically when needed. A new research paper argues this implicit layer is actually more critical for long-running agents and shouldn't depend on agent initiative.

The Design Gap#

Existing coding agents ship with only one memory tier: deliberate storage that agents choose to write and read. The research team observed that agents almost never use voluntary memory systems even when primed with information—in a real coding task spanning 114 turns, agents performed zero memory operations. This mirrors how humans rely less on explicit notes and more on cues that trigger knowledge automatically.

A Cue-Anchored Approach#

The proposed architecture moves memory delivery into the harness layer—the environment controlling the agent—rather than leaving it to agent choice. The system uses a cue-anchored model where memories carry first-class trigger conditions. These triggers span a composable vocabulary of activation cues: path-based (certain files), symbol-based (function names), semantic (related concepts), event-based (actions taken), and temporal (time windows). The harness evaluates these conditions deterministically and injects relevant facts without agent overhead.

Validation on Real Work#

Testing on an actual coding task revealed stark contrasts. Agents left voluntary memory untouched (zero operations across 114 turns). In contrast, deterministic harness-injected memories arrived reliably in every seeded run with zero false positives. Even more revealing: when a compaction operation (summary generation) happened, voluntary memories in conversation vanished and stayed absent in 106 of 108 subsequent compactions. Meanwhile, the same facts stored in the harness persisted intact through all 138 compact-resume cycles, never requiring agent re-retrieval.

Delivery Over Storage#

The core insight is that reliable memory for long-running agents isn't about better storage—it's about reliable delivery. The findings suggest the most effective memory channel is one agents never have to think about consciously. By anchoring memories to environmental cues and moving their injection to the harness layer, the architecture removes the friction that causes agents to drift away from memory tools entirely.