Finding
Recuris, arXiv:2608.24876, submitted 2026-08-25, evolves agent memory rather than weights or a monolithic prompt. A frozen agent uses Working Memory to track current task state and select from Experiential Skill Memory. Structured execution traces localize failures to a specific memory component, then a fixed meta agent patches only that component. A deterministic paired held out gate decides whether the patch survives.
The originating team reports improvement in 35 of 37 completed model benchmark pairs across four long horizon benchmarks and ten models. Reported gains include plus 17.8 points for GPT 5.6 Sol on tau bench, plus 15.6 for Claude Opus 5, and up to plus 32.2 points on the longest tasks. Common long horizon failure modes fall by up to 80% in their evaluation.
Code was released under Apache 2.0 at Gen-Verse/Recuris. Pin reproduction work to commit f54c9dabfa370c0da495ddabe8ccbe8702b3eae7 or a reviewed successor rather than tracking moving main.
Evidence class: originating team report with released code and frozen evaluation splits, not independently reproduced by RuV.
Opportunity score
Weighted Opportunity Score: 4.66 of 5.
Why this is not duplicate work
Dream Machine PR #30 already implements trace driven harness optimization with held out gates. Core Memory already has state first projection, and PR core-memory#17 adds persistent verifier signals. The missing reusable primitive is failure localized evolution across distinct working state and experiential skill memory instead of treating the harness as one patch surface.
Proposed RuV experiment
Define a memory evolution adapter with four explicit components:
- current verified working state from Core Memory
- experiential skill content
- skill applicability and routing rules
- validation and promotion criteria
Each execution trace records current state, selected skill, action, observation, outcome, and verifier evidence. Failure attribution may nominate exactly one memory component per standard mutation. Dream Machine then generates a candidate patch and runs the existing held out promotion gate from PR #30.
Benchmark
First reproduce one upstream benchmark with the upstream implementation and pinned environment. Then compare three RuV conditions under identical model, task split, seed, budget, and evaluator:
- base agent
- base agent plus static RuV skill memory
- base agent plus failure localized RuV memory evolution
Report success, horizon bucket, tool errors, stale state errors, wrong skill invocation, tokens, wall time, API cost, accepted and rejected mutations, variance, and failure category.
Falsification
Reject the mechanism as a RuV production primitive if gains disappear under held out tasks, if one memory component accumulates most changes without attributable benefit, if static curated memory matches the evolving version within variance, or if evolution increases cost more than the task success value justifies.
Cross stack mapping
Core Memory supplies current state and provenance. RuVector indexes experiential skills. Dream Machine generates bounded mutations. MetaHarness runs independent evaluation. Ruflo supplies long horizon executions. RVM prevents memory evolution from changing runtime authority. Persistent verifier signals from core-memory#17 can annotate accepted and rejected experience.
Acceptance
Do not promote based on upstream numbers. Require at least 5 absolute percentage points held out task success improvement over the stronger of base and static memory, no authority expansion, no evaluator mutation, no protected slice regression above 2 points, and full lineage for every accepted memory change.
Finding
Recuris, arXiv:2608.24876, submitted 2026-08-25, evolves agent memory rather than weights or a monolithic prompt. A frozen agent uses Working Memory to track current task state and select from Experiential Skill Memory. Structured execution traces localize failures to a specific memory component, then a fixed meta agent patches only that component. A deterministic paired held out gate decides whether the patch survives.
The originating team reports improvement in 35 of 37 completed model benchmark pairs across four long horizon benchmarks and ten models. Reported gains include plus 17.8 points for GPT 5.6 Sol on tau bench, plus 15.6 for Claude Opus 5, and up to plus 32.2 points on the longest tasks. Common long horizon failure modes fall by up to 80% in their evaluation.
Code was released under Apache 2.0 at Gen-Verse/Recuris. Pin reproduction work to commit
f54c9dabfa370c0da495ddabe8ccbe8702b3eae7or a reviewed successor rather than tracking moving main.Evidence class: originating team report with released code and frozen evaluation splits, not independently reproduced by RuV.
Opportunity score
Weighted Opportunity Score: 4.66 of 5.
Why this is not duplicate work
Dream Machine PR #30 already implements trace driven harness optimization with held out gates. Core Memory already has state first projection, and PR core-memory#17 adds persistent verifier signals. The missing reusable primitive is failure localized evolution across distinct working state and experiential skill memory instead of treating the harness as one patch surface.
Proposed RuV experiment
Define a memory evolution adapter with four explicit components:
Each execution trace records current state, selected skill, action, observation, outcome, and verifier evidence. Failure attribution may nominate exactly one memory component per standard mutation. Dream Machine then generates a candidate patch and runs the existing held out promotion gate from PR #30.
Benchmark
First reproduce one upstream benchmark with the upstream implementation and pinned environment. Then compare three RuV conditions under identical model, task split, seed, budget, and evaluator:
Report success, horizon bucket, tool errors, stale state errors, wrong skill invocation, tokens, wall time, API cost, accepted and rejected mutations, variance, and failure category.
Falsification
Reject the mechanism as a RuV production primitive if gains disappear under held out tasks, if one memory component accumulates most changes without attributable benefit, if static curated memory matches the evolving version within variance, or if evolution increases cost more than the task success value justifies.
Cross stack mapping
Core Memory supplies current state and provenance. RuVector indexes experiential skills. Dream Machine generates bounded mutations. MetaHarness runs independent evaluation. Ruflo supplies long horizon executions. RVM prevents memory evolution from changing runtime authority. Persistent verifier signals from core-memory#17 can annotate accepted and rejected experience.
Acceptance
Do not promote based on upstream numbers. Require at least 5 absolute percentage points held out task success improvement over the stronger of base and static memory, no authority expansion, no evaluator mutation, no protected slice regression above 2 points, and full lineage for every accepted memory change.