flâneur

Memory in the wild: how we use Context Engine on our own code | Applied Compute

appliedcompute.com · 1,916 words · saved by 1 readers

Last week⌝ we introduced Context Engine, Applied Compute’s system for remembering, refining, and retrieving enterprise context to build continual learning agents. This post is about what happened when we pointed it at ourselves. Although the results are early, they show promising gains from using context derived from prior traces. Notably, using a Contextbase built over our own coding sessions, we were able to roughly 2x the rate of retrieving memories critical to a coding agent’s task through continual learning in production. Additionally, on a small curated set of tasks where we found clear indication of memories being critical, agents were able to successfully use injected memories from the Contextbase to outperform the no-memory baseline. For the past few months, we've logged every coding agent interaction at Applied Compute across Cursor, Claude Code, and Codex, and routed those traces into Applied Compute Logs (ACL), a single log of how we actually build software. We then used ou

Last week ⌝ we introduced Context Engine, Applied Compute’s system for remembering, refining, and retrieving enterprise context to build continual learning agents. This post is about what happened when we pointed it at ourselves. Although the results are early, they show promising gains from using context derived from prior traces. Notably, using a Contextbase built over our own coding sessions, we were able to roughly 2x the rate of retrieving memories critical to a coding agent’s task through continual learning in production. Additionally, on a small curated set of tasks where we found clear

saved by

related reading