After the Prototype - agents, applied AI, and models put to work

Every engineer on a team now runs a coding agent, and every agent builds its own picture of the codebase. The shared layer on top - decisions, conventions, the rules in CLAUDE.md or AGENTS.md - goes stale the moment code changes under it, and an agent acts on an outdated fact with the same confidence as a true one.

We build MEX (open source, github.com/mex-memory/mex): shared project memory for engineers and their coding agents that lives in the repo and syncs through Git. This talk presents what we measured. We replayed 367 real commits of the Hono web framework and checked which tools still served outdated facts as current: memory tools with no link to code served 90-100% of outdated facts as true; MEX served 20%. The detection is deterministic - each stored fact links to a code-graph node and is flagged when that code changes, no LLM in the check. The cost is false warnings (56.9% fresh retention with plain flagging; 89.2% after an LLM re-review step). Underneath, the code graph put the fix-relevant code in the top five for 65.9% of 656 SWE-Explore issues against 30.2% for keyword search, with 24 of 24 incremental updates identical to a clean rebuild. I will also cover why we chose the repo as the home and Git as the transport for shared memory, and what that boundary means for review and approval.

Takeaways: measured evidence of how fast agent-facing project knowledge goes stale, and a deterministic, code-linked way to detect it that works without an LLM - plus honest numbers on the false-warning cost and how to recover from it.

Audience: engineers and engineering leads running coding agents on shared codebases, and anyone evaluating memory or context tooling for agent workflows.

Session format: 30-minute talk (no demo).

Bio: Archit Dhamija is the founder of MEX (mexmemory.com), open-source shared memory for coding agents.

Full benchmark with setup and limits: https://mexmemory.com/blog/memory-that-goes-stale