Test with confidence, code with clarity

• Persistent memory is being sold as the next harness layer. A 2026 wave of papers and tools (arxiv 2606.30306 survey, arxiv 2609.05510 operational report, arxiv 2606.12329 PROJECTMEM, arxiv 2604.27283 abstention-aware retrieval, the NeurIPS 2026 selective-retrieval paper) proposes cross-session memory as the fix for coding agents that forget what they already learned. The pitch is real and the layer is real.
• Memory without validation propagates past wrong fixes. Naive similarity retrieval shipped a 75% false-positive rate on one hard-negative benchmark. Agents re-apply the fix that passed last time to the case it does not actually fit. The memory is faithfully doing its job and the suite is still the authority that catches the propagation.
• Selective retrieval helps because it is a test-discipline move. The NeurIPS result (29 of 94 resolved with validated-episode retrieval versus 19 of 94 with none) is what works. The mechanism is a validation gate the memory writes through, not the memory itself. Memory earns its keep when the suite told it the past fix was correct for this family of cases.
• When memory and the suite disagree, the suite wins. A memory system is a cache of past readings that can be stale, wrong, or irrelevant to this change. The test the agent reads before every write is the primary signal; the memory is a hint. Any workflow that treats the memory as authoritative is selling the agent its own homework as the verifier.

Read the argument: The Agent's Memory Is Not a Test Suite