AuthorsZhaochen Yu, Yingcheng Wu, Zhenfei Yin et al.
TL;DR
Recuris couples verified working memory with experiential skills and a recursive Meta-Agent, adding up to +32.2 points on the longest long-horizon tasks.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
As interaction histories grow, agents retrieve against noisy dialogue and lose track of unresolved goals, causing misaligned skill use and stalled execution.
On π2-Retail, the base agent omits all required database writes in 42% of episodes and achieves only 55.7% required-write recall, leaving tasks apparently complete but environments unchanged.
HOW IT WORKS
Recuris represents Skill Memory as Mπ = (Eπ, Wπ, ππ, Cπ), coupling Experiential Memory Eπ with Working Memory specification Wπ, invocation policy ππ, and checker set Cπ.
Think of Working Memory π€π‘ as RAM tracking current goals, while Experiential Memory Eπ is a disk of reusable skills that Recuris selectively loads based on the verified task state.
This EMβWM coupling plus a fixed Meta-Agent turns execution into structured evidence, enabling component-level repairs to Mπ that a plain context window or static skill library cannot provide.
DIAGRAM
This diagram shows how Recuris uses verified working memory to drive state-grounded skill invocation and evidence-grounded state updates at each step.
DIAGRAM
This diagram shows how Recuris localizes failures from structured traces and applies validation-gated, component-specific patches to Skill Memory Mπ.
PROCESS
Structured Working State
Recuris initializes working state w0 under Working Memory specification Wπ, creating goal entries with status, evidence, and blockers to expose pending and done goals.
State-Grounded Skill Invocation
At each execution event ππ‘, Recuris uses invocation policy ππ to select skills Eπ‘ from Experiential Memory Eπ based on the current verified state π€π‘.
Evidence-Grounded State Update
After the LLM acts and Env returns ππ‘, Recuris proposes eπ€π‘+1 via Wπ, evaluates it with checker set Cπ, and commits only supported changes through kernel K.
Bounded Recursive Evolution
Across tasks, a fixed Meta-Agent reads structured traces Ξπ, runs trace-based failure localization, applies component-specific patching to Mπ, and admits updates via a validation gate.
KEY CONTRIBUTIONS
ExperientialβWorking Memory Coupling
Recuris couples Experiential Memory Eπ with verified Working Memory Wπ, enabling state-grounded skill invocation that raises required-write recall on π2-Retail from 55.7% to 82.4%.
Structured Evidence for Failure Localization
Recuris records structured traces Ξπ linking states, skills, actions, and observations, boosting component fault recall from 13.0% with outcomes to 64.8% with Ξπ.
Bounded Recursive Skill Memory Evolution
Recuris evolves Skill Memory Mπ via a fixed Meta-Agent and validation gate, yielding held-out gains up to +17.44 points over M0 and transferring +17.8 points to GPT-5.6 Sol on π2-Retail.
RESULTS
Task success on Ο2-Retail (doubao-2.0-pro)
81.4%
+23.3 points over base agent 58.1%
Task success on SkillFlow (Qwen3.6-27B)
58.7%
+16.6 points over base agent 42.2%
Task success on Ο2-Retail (Claude Opus 5)
87.9%
+15.6 points over base agent 72.4%
Failure-mode reduction (hallucination completion)
14%
β86% failure rate relative to agent alone = 100
Recuris is evaluated on π2-Retail, π2-Airline, SkillFlow, and Terminal-Bench 2.1, covering tool-use dialogue, lifelong skill reuse, and terminal work. The gains up to +32.2 points on the longest tasks and up to 86% reduction in common long-horizon failures show that Recuris turns accumulated experience into reliably improved long-horizon behavior.
BENCHMARK
Recuris is evaluated on π2-Retail, π2-Airline, SkillFlow, and Terminal-Bench 2.1, covering tool-use dialogue, lifelong skill reuse, and terminal work. The gains up to +32.2 points on the longest tasks and up to 86% reduction in common long-horizon failures show that Recuris turns accumulated experience into reliably improved long-horizon behavior.
BENCHMARK
Task success (%) on Ο2-Retail for the base agent and Recuris across several models.
KEY INSIGHT
On π2-Retail, adding experiential skills alone yields only +2.0 points, while adding a verified working state yields +23.9 points and coupling both reaches +25.4.
This is counterintuitive because many systems focus on richer skill libraries, but Recuris shows that when and why skills are invoked, via working memory, matters far more than skill content alone.
WHY IT MATTERS
Recuris unlocks long-horizon agents whose memory-control layer can recursively evolve while the base LLM and outer harness remain frozen, enabling safe, interpretable RSI.
Builders can now ship agents that continuously convert structured execution traces into targeted Skill Memory updates, improving tool use, goal tracking, and failure recovery across tasks and even across models.
Nikhil Verma
Β· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M β 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
Β· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRAβs EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
Β· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checkingβ¦