Hippo preprint proposes training-free agent memory built by comparing attempts

A preprint submitted to arXiv on 17 October 2026 proposes Hippo, a memory framework intended to let LLM agents improve from their own experience without updating model parameters. The authors frame the work around recursive self-improvement and position external memory as the adaptation channel. Training-free here means no weight updates; it does not mean the method adds no computation, and the abstract gives no figure for the overhead involved. Source

The case for the design rests on two problems in learning from experience. A single attempt is a poor teacher, the authors argue: flawed reasoning can produce success, and poor execution can obscure a sound strategy. Accumulating lessons without consolidation has its own failure mode. The authors report that their accumulation baseline without consolidation improved at first but eventually performed below a no-memory baseline. Hippo’s response is to compare attempts at the same task, identify decisions that look reusable, and refine the conditions under which they apply. Source

The hippocampal reference sits mainly in how memory is formed rather than in any claim to model biology. The authors describe the method as inspired by selective hippocampal encoding and replay, learning from comparisons when shortcomings are diagnosed. New experience is checked against existing memory, with overlaps merged and conflicts resolved through explicit applicability conditions. The stated benefit is that earlier guidance can be revised rather than appended, so lessons do not pile up as separate entries. Source

The reported evaluation spans SWE-bench Verified, WebArena and Mind2Web, with performance gains in all three and substantially fewer retained experiences than the accumulation baseline. The authors also state that the gains hold across evaluated base models of varying capability, which supports their argument that memory-based adaptation remains useful for stronger models. These are the authors’ results as summarised in the abstract; no independent reproduction or peer review is available. Source

For teams running agents, the interesting claim is about maintenance rather than a single benchmark jump. Consolidation is meant to keep a growing memory usable, and the accumulation baseline is presented as evidence that growth alone degrades. Whether Hippo’s merge-and-condition approach is what prevents that degradation, and at what cost in comparison and conflict-resolution work, is not shown in the abstract. Source

What the abstract leaves open Source

No figures appear in the abstract: no per-benchmark scores, no measured size of the gains and no count for the reduced experiences. Without those numbers, the savings from a smaller memory cannot be weighed against the work of building it. The list of base models is also absent, so it is unclear how far the result carries beyond the three evaluated environments. The listing retrieved shows no code link or licence, so reimplementation cost and third-party results remain unknown. Source

Produced through the Singularity Logs automated editorial workflow, with separate factual and editorial AI checks. Source links identify the evidence used.