Reglos splits personal-agent memory into two loops: memory grows with each chat while a coding agent rewrites the harness

Reglos divides a personal memory agent's improvement into two layers: a per-user memory record that grows with conversation, and the harness code that decides which memories the model reads. Earlier memory systems updated what they stored, but the harness was usually written by hand and revised between releases; handing it to a coding agent that evolves against a benchmark leaves open whether a change also helps users whose failures nobody inspected while writing it. With Qwen3.8-27B writing memory for every system and answering, Reglos reaches 61.3% on the official PersonaMem-v2 test set, 7.5 points above the strongest compared system and 12.0 points above the same model reading the full history.

Memory is kept as an append-only record: each fact carries a validity interval and a time learned, a new preference closes the old fact's validity without erasing it, and a request to forget expires matching facts and leaves a tombstone. The harness evolves through recursive self-improvement: a coding agent reads raw failure traces and writes patches, while a critic from a second model family clusters the same failures independently and rejects patches keyed to the dataset's option templates or labels. After the freeze, a fixed evaluator requires a candidate to gain on each of two persona-disjoint validation halves and to cost the second reader, GPT-5.6 Luna, at most one point; four accepted steps lifted the harness from 54.2% to 60.7%.

The demonstrated scope is personal conversational memory: on PersonaMem-v2, LongMemEval-S, LoCoMo and BEAM, with Qwen3.8-27B writing memory and either Qwen3.8-27B or GPT-5.6 Luna answering, compared with systems including Mem0, MemOS and EverOS; harness evolution ran only on the PersonaMem-v2 validation split, and the test set was scored once after the freeze.

In the development audit about half of the remaining errors needed facts never stated in the conversation, putting the ceiling for a perfect reader near 80%; preferences belonging to other people stay below chance for every system, extracted as the user's own and trusted by later layers. The LoCoMo result rests on passing short histories to the reader in full, and on BEAM the systems sit within a few points of each other; the published best result it compares with uses a different reader, GPT-4.1-mini.