acceptodds
Under review as a conference paper at ICLR 2027

OpenKAGE: Open-Ended Lifelong Learning in the Open World via Knowledge-Sharing Agent-Group Evolution

Abstract

Open-ended lifelong learning requires agents to accumulate and reuse knowledge while autonomously selecting goals. Existing research often pursues lifelong learning under predefined tasks, curricula, or objectives, and largely entrusts self-evolution to a single agent. We propose OpenKAGE, a training-free framework that realizes this vision through group-level knowledge sharing in open-world environments: (1) each agent maintains a local knowledge graph with two layers—a private memory layer for situated experience and a shareable experience layer holding transferable knowledge as textual records and executable skills; (2) a semantic aggregator periodically consolidates the experience layers into a shared graph, which agents access through intent-aware subgraph retrieval, while memory-layer records remain local. In 30-game-day Minecraft evaluations with two backbone models, OpenKAGE achieves obsidian acquisition and Nether entry, milestones not reached by a matched planner-only baseline, while reducing mean deaths. A cross-group, cross-backbone transfer experiment further suggests that accumulated knowledge can benefit newly initialized agents. Together, these results demonstrate sustained autonomous learning and effective knowledge reuse within the evaluated horizon, supporting OpenKAGE as a practical framework for open-ended lifelong learning.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Open-Ended Lifelong Learning, Knowledge-Sharing, Agent-Group Evolution, Minecraft

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_nqUpOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary
  1. OpenKage is a framework for consolidating knowledge or skills learned by the agents when running in Minecraft without model weight updates to improve the agent's performance at test time.

  2. Its main contributions are 1) proposing a framework for knowledge sharing and transfer during test time to improve the agent's performance, and 2) using a specific group-level aggregator to consolidate textual experience into a shared knowledge graph to avoid the conflicts of context-specific skills. This is achieved by using a reflector to separate situated memories from transferable experience. The shared knowledge is sent back to the agent through a graph-based intent-aware retriever. 3) shows the cross-model and cross-agent transfer ability of learned skills

  3. Validation: the paper mainly evaluates its approach against a planner-only baseline on Minecraft using 2 models. The result is that OpenKage gets better exploration while reducing death rates. The results change across different models.

Critical strengths

The most promising part of the paper is the separation of private history and shareable skills across agents. Nevertheless, more ablations should be added to demonstrate that the proposed mechanism is what really contributes to the success.

Critical weaknesses
  1. The novelty of the proposed framework is a main concern. For instance, works such as G-memory [1] and [2] are already exploring similar approaches in terms of skill learning and memory consolidation. However, the paper fails to discuss the broader related works and does not differentiate itself from the literature well.

  2. Beyond the novelty issue, the paper has a limited experimental setting. The current experiment only operates over Minecraft and does not propose fundamental algorithmic-level improvement in terms of training. If the main focus is on agent harness, then the paper should compare against a few more agent harness baselines to demonstrate the practical applicability of the proposed harness for multiagent systems. The ablation studies are insufficient to support the claims proposed in the paper. If the goal is to show that cross-trial or cross-agent experience sharing enables better test-time learning, then only testing on Minecraft is not enough.

[1] Zhang, Guibin, et al. "G-memory: Tracing hierarchical memory for multi-agent systems." Advances in Neural Information Processing Systems 38 (2026): 12988-13018.

[2] Wu, Shanglin, et al. "Scaling Teams or Scaling Time? Memory Enabled Lifelong Learning in LLM Multi-Agent Systems." arXiv preprint arXiv:2604.03295 (2026).

Questions for authors
  1. How would you position your work against the two related works I listed above?
  2. How to ensure the reproducibility of the performance gain from Minecraft across diverse tasks if we want to apply OpenKage to them? Have you tested on more benchmarks beyond Minecraft?
  3. How to control the randomness introduced in the experiment? What is the uncertainty in running the long-horizon tasks to demonstrate the usefulness of the proposed harness?
  4. Have you compared your framework against established single-agent skill consolidation and learning frameworks such as Voyager [1] or other general harnesses for multiagent skill consolidation over long-horizon tasks?

[1] Wang, Guanzhi, et al. "Voyager: An open-ended embodied agent with large language models." arXiv preprint arXiv:2305.16291 (2023).

8 comments
AuthorsOpenReview ↗

Experiment tables of the revised paper

We had been working on a larger re-run since October, and the revised paper reports it: every cell below has 10 independent group runs of 3 agents (30 agents per cell), on world seed 7 and 30 game days. The re-run added the Voyager and Echo comparisons, three further backbones and the Raw ablation condition. Below are the three experiment tables of the main text (Tables 1–3) so that all of them can be read in one place; per-run values are in Tables 13 and 15 of the appendix, and everything that changed relative to the initial submission is marked in blue in the PDF.

Metrics. Deaths, Repeat (deaths repeating an earlier one of the same agent, same cause within 16 blocks) and XP/life are per-agent means; lower is better for Deaths and Repeat, higher for XP/life. Tool era gives the highest tier reached (dmd. = diamond), the mean day of first attainment and the number of agents out of 30. Obsidian and Nether give the mean first-attainment day and the number of agents out of 30; "—" means that no agent attained it.

Table 1. Comparison with baselines on DeepSeek-V4.1-Flash.

MethodDeathsRepeatXP/lifeTool eraObsidianNether
Voyager228.9203.11.0dmd. (3.0 day, 1/30)——
Echo73.653.14.2dmd. (15.8 day, 26/30)——
Planner-only27.915.114.0dmd. (15.2 day, 16/30)——
OpenKAGE16.27.418.2dmd. (13.7 day, 24/30)22.8 day (6/30)23.8 day (4/30)

Table 2. OpenKAGE and its planner-only control across backbones.

BackboneMethodDeathsRepeatXP/lifeTool eraObsidianNether
DS-V4.1FPlanner-only27.915.114.0dmd. (15.2 day, 16/30)——
OpenKAGE16.27.418.2dmd. (13.7 day, 24/30)22.8 day (6/30)23.8 day (4/30)
GLM-5.3Planner-only33.417.48.1dmd. (15.2 day, 21/30)——
OpenKAGE28.616.924.1dmd. (13.8 day, 28/30)23.9 day (10/30)27.0 day (2/30)
hy4-previewPlanner-only47.028.88.2dmd. (16.3 day, 29/30)——
OpenKAGE15.68.332.4dmd. (18.8 day, 13/30)28.0 day (1/30)29.0 day (1/30)
Kimi-K3Planner-only64.948.77.0dmd. (21.7 day, 15/30)——
OpenKAGE42.127.06.5dmd. (19.1 day, 10/30)22.0 day (2/30)—

Table 3. Ablations on DeepSeek-V4.1-Flash. Conditions are named by what they remove: −S sharing, −R subgraph retrieval, −M the memory layer, −E the experience layer, B all graphs (the planner-only control); Raw keeps every component but shares experience records verbatim (Appendix B.5). Full is the OpenKAGE row of Table 1.

ConditionDeathsRepeatXP/lifeTool eraObsidianNether
Full16.27.418.2dmd. (13.7 day, 24/30)22.8 day (6/30)23.8 day (4/30)
Raw17.610.821.6dmd. (13.0 day, 14/30)——
−S17.48.116.1dmd. (13.3 day, 21/30)21.5 day (2/30)21.0 day (1/30)
−R15.67.413.3dmd. (13.3 day, 12/30)——
−M11.64.137.7dmd. (17.9 day, 19/30)——
−E14.57.421.7dmd. (14.0 day, 15/30)——
B27.915.114.0dmd. (15.2 day, 16/30)——

Thank you for the review and the two references.

Weakness 1 / Q1. Positioning against G-Memory [1] and "Scaling Teams or Scaling Time" (LLMA-Mem) [2]. We read both. Both are now in the Introduction and Related Work, and Table 11 (Appendix A.10) has a row for each.

  • G-Memory manages multi-agent interaction through a three-tier graph hierarchy (insight, query, interaction); it retrieves related past task queries with their insights and trajectories, filtered by role. That hierarchy is keyed by the task queries the system receives. In our setting agents set their own goals, so the query level has no counterpart and the design transfers only with changes. It is evaluated on benchmarks with assigned tasks.
  • LLMA-Mem [2] keeps episodic, procedural and transactive memory under local, shared or hybrid topologies, on assigned tasks. Its hybrid topology keeps episodes local and shares consolidated procedures. That design choice is shared with our work, and the paper credits it. OpenKAGE differs in four respects: the boundary is drawn by the kind of knowledge, enforced by access (the aggregator receives only experience-layer records); the shared object is a graph whose applicability conditions travel with the text, while executable skills are shared verbatim and remain maintained by their author; retrieval returns subgraphs; and agents set their own goals.

G-Memory and LLMA-Mem are outside our comparison set: their memory structures build on task queries or assigned tasks, so a port would be our own re-implementation. Our two nearest controlled stand-ins are Raw (shared memory without consolidation) and −S (private structured memory). A head-to-head comparison with either system is future work.

Weakness 2: limited setting, training-level algorithmic change. OpenKAGE is a framework that drives lifelong learning and collective evolution of agents: capabilities grow through structured memory, consolidation and cross-agent sharing of experience, so the algorithmic contribution of this paper lies at the level of knowledge and memory mechanisms; carrying learning further into model parameters is a complementary direction for future work. Evaluation is in Minecraft, as the Limitations state; other environments and benchmarks are future work. Within Minecraft, our evidence is: four backbones, 10 runs per cell, three baselines (planner-only, Voyager, Echo), seven ablation conditions with 10 runs each, and one cross-backbone transfer experiment.

Ablations insufficient. Partly agreed. The ablations do reflect that the components matter jointly: every ablated condition falls short of Full on the deepest milestones (obsidian in 0 to 2 of 30 agents vs 6) while keeping deaths well below planner-only (11.6 to 17.6 vs 27.9 per agent). Apportioning the effect to individual components needs more runs: deep milestones are rare even under Full, and −S, −R, −E and Raw stay close to Full in Deaths, Repeat and XP/life relative to the run-to-run spread. Ablation claims that rested on single runs are removed. The best-supported statement remains the combined stack vs planner-only: Deaths 16.2 vs 27.9 per agent, repeated deaths 7.4 vs 15.1, and obsidian in 13/40 vs 0/40 runs pooled over four backbones.

Q2. Reproducibility across tasks / other benchmarks. Our tests are in Minecraft, on world seed 7. Within Minecraft, reproducibility of the gain is as follows: the survival effect repeats on four backbones (mean Deaths lower on all four). The exploration effect appears on every backbone (obsidian in 4/10, 6/10, 1/10, 2/10 OpenKAGE runs vs 0/10 for every planner-only cell; its size differs by backbone). For other domains the framework reserves a domain-extension interface: each task domain is a plug-in package implementing a task-domain protocol, while the framework core is written once against that protocol. The code for the framework, baselines and metrics is released; we leave instantiating this interface on other tasks and benchmarks to future work.

Q3. Randomness and uncertainty in long-horizon runs. Appendix B.1 ("Sources of randomness"): world seed, spawn point, difficulty, personas, prompts and hyperparameters are fixed; LLM sampling (API-internal seeds), the edge sampling of relational expansion (unseeded generator) and game dynamics (mob spawns, combat, asynchronous timing) vary between runs. A run's number records launch order only. We therefore report the run-to-run spread directly: 10 runs per cell, all runs listed in Tables 13 and 15, Fig. 9 showing each run, the run as the unit of counting (agents of a run share one graph). Example: on the default backbone, deaths per agent in the ten OpenKAGE runs range from 9.7 to 26.0, against 19.0 to 41.3 for planner-only, and every OpenKAGE run is below the planner-only median. On long horizons: all main runs are 30 game days; the 60-day and group-size runs (B.6) are single runs and labelled as such.

Q4. Comparison with Voyager and other harnesses. Voyager: yes. We re-implemented Voyager (curriculum, iterative code generation with execution feedback, self-verification, per-agent skill library) under the same server, world, Mineflayer API and horizon, with its high-level control functions removed and three isolated agents per group. Voyager's high-level control functions are hand-written capability primitives (mineBlock, craftItem, smeltItem, killMob) that hand the agent, at step one, composite abilities our protocol treats as the outcome to be learned. Keeping them would, under our protocol, give the baseline a large advantage unrelated to the mechanism being compared, akin to playing with x-ray vision and cheats enabled: mineBlock scans every matching block within 32 blocks through solid rock in one sweep and then handles pathfinding, digging and item pickup end to end, so an agent can mine before it has learned anything. The comparison would measure pre-supplied primitives rather than accumulated knowledge (A.3, A.10). Both external baselines therefore act through the same bare Mineflayer API as OpenKAGE, so that what differs is the mechanism turning the goal into tasks and experience (B.3). Because the hard-coded first Voyager task can fail repeatedly under this API. Echo (Li et al., 2026): ported mechanism by mechanism, because its released code lacks its runtime. On the default backbone, per agent: Deaths 228.9 (Voyager), 73.6 (Echo), 27.9 (planner-only) vs 16.2 (OpenKAGE); diamond-tier tools 1/30, 26/30, 16/30 vs 24/30 agents; obsidian and the Nether appear only under OpenKAGE. Both external baselines are mechanism-level re-implementations under one protocol.

Presentation. We rewrote the Limitations so that the scope is explicit (Minecraft, one world seed in the main experiments, 30-day horizon) and added the appendix tables listed above; the paper presents OpenKAGE as a framework that drives lifelong learning and collective evolution of agents. If specific passages remain hard to follow, please point us to them and we will revise.

Reviewer_HejuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

OpenKAGE combines private agent memories, shared experience graphs, and reusable executable skills for autonomous Minecraft exploration. Agents reflect locally, periodically consolidate transferable knowledge, and retrieve relevant subgraphs. Experiments examine exploration, survival, component ablations, and knowledge transfer between groups and backbones.

Critical strengths
  1. The separation of private memories, shared experience, and executable skills gives a concrete design for reusable group knowledge. Its value is the complete accumulation-and-reuse workflow, which others can inspect and build on.
  2. The main result repeats across three groups on each of two backbones: every full-system group acquires obsidian, while no planner-only group does. The no-sharing ablation and cross-backbone transfer add useful, though preliminary, evidence for knowledge reuse.
Critical weaknesses
  1. The strongest repeated comparison removes memory, reflection, skills, and sharing together. It establishes the full pipeline's benefit, but leaves the added value of group knowledge consolidation over local learning and basic record sharing less clear.
  2. The no-sharing control appropriately retains local capabilities, but component and transfer studies mostly use one group on one world seed. These more diagnostic experiments need stronger support to distinguish reliable benefits of knowledge sharing from variation between groups.
  3. The reliability of shared knowledge is not directly audited: skill-call completion does not guarantee goal completion, and contradiction handling is prompt-driven. Comparison with established memory-equipped systems such as Voyager would also help position the contribution.
  4. Paused game time leaves inference costs outside the comparison. Reporting reflection, aggregation, and execution costs would make the practical trade-off clearer; the current results establish progress within a game horizon rather than computational efficiency.
  5. Minor corrections: Section 3.3 should not describe the Nether rate of -R as below Full: 1/3 equals 3/9. Clarify Table 14's “obsid.” tool-era entry and remove the repeated opening explanation in B.5.
Questions for authors
  1. What evidence separates the benefit of semantic consolidation from directly sharing the same experience records with local memory and skills preserved?
  2. Can the existing results report uncertainty at the group level, and distinguish repeated findings from exploratory single-group observations?
  3. In a sample of shared records and skill calls, how often are applicability conditions preserved and intended goals actually achieved?
  4. What are the measured token, call, and runtime costs of the main system components?
5 comments

Thank you. Your minor corrections are right, and we answer the four questions in order.

Weakness 1 / Q1. Evidence separating semantic consolidation from directly sharing the same records with local memory and skills preserved. We added exactly this control, Raw (Table 3, Appendix B.5, Table 15): private graphs, memory layer, retrieval budget and skill sharing as in Full; only aggregation differs. Records from the agents' experience layers are appended to the shared graph verbatim with window and agent prefixes, with merging, deduplication and conflict handling bypassed and the aggregation agent left idle. Ten runs each: Deaths 17.6 vs 16.2, Repeat 10.8 vs 7.4, XP/life 21.6 vs 18.2, weakest-agent XP/life 5.9 vs 7.9; obsidian 0/10 vs 4/10 runs, Nether 0/10 vs 3/10. The means of Raw and Full lie within each other's run-to-run spread, and Appendix B.5 and the Limitations describe the comparison at this level. The sharing channel is used equally in both (62% vs 60% of skill calls invoke another agent's skill; 94% and 95% run without error), so the comparison isolates consolidation inside an active sharing channel. Two descriptive observations: the final shared graph holds a median of 108 concepts under Full vs 348 under Raw; and Raw's gap appears early: 16/30 Raw agents never obtain a diamond pickaxe (6/30 under Full), while among agents who do, the share proposing an obsidian goal is similar (8/14 vs 18/24) and fewer obtain an iron bucket (2/14 vs 13/24). The paper reports this gap, notes that two Raw agents held the complete pouring procedure and still failed, and offers one possible attribution: verbatim records keep events one by one, so the actionable rule takes up little of the retrieved block.

Weakness 2 / Q2. Group-level uncertainty; repeated findings vs exploratory observations. The unit is now the run. The revised paper separates three tiers:

  • Repeated, 10 runs per cell: Tables 1–3; every run is listed in Table 13 and Table 15, and Fig. 9 plots each run.
  • Three runs per group: the transfer experiment (Fig. 6; nine-agent means, so descriptive).
  • Single run, labelled supplementary: group-size and 60-day runs (B.6).

Cell means and run counts for the headline contrasts are in the General response and in our reply to ao7v (comment 2). The component ablations run on a single seed; at 10 runs they are consistent with every component mattering, and apportioning the effect among components needs more runs.

Weakness 3 / Q3. Is shared knowledge reliable? How often are applicability conditions preserved and goals achieved? Skill-call completion and goal completion are different quantities; the paper defines success as "returned without error" and states that a call that returned without error need not have reached its goal (B.4). What the paper reports for shared skills is that a skill fails about as often in the hands of another agent as in its author's (5.6% of 3,526 calls for teammates' skills vs 6.9% of 2,206 for the agent's own on the default backbone), and that failed skills are revised by their authors (the failure rate of a skill's calls falls from 31.3% to 7.3% after revision, with the regression-to-the-mean caveat given there). Contradiction handling is prompt-driven, as you note. Measuring how often applicability conditions are preserved and goals are achieved needs a human-labelled gold standard, which we leave to future work.

Comparison with Voyager. Done (Table 1, B.3): Deaths 228.9 vs 16.2 per agent; diamond-tier tools 1/30 vs 24/30; obsidian appears only under OpenKAGE, with Voyager and Echo at 0.

AuthorsOpenReview ↗

Experiment tables of the revised paper

We had been working on a larger re-run since October, and the revised paper reports it: every cell below has 10 independent group runs of 3 agents (30 agents per cell), on world seed 7 and 30 game days. The re-run added the Voyager and Echo comparisons, three further backbones and the Raw ablation condition. Below are the three experiment tables of the main text (Tables 1–3) so that all of them can be read in one place; per-run values are in Tables 13 and 15 of the appendix, and everything that changed relative to the initial submission is marked in blue in the PDF.

Metrics. Deaths, Repeat (deaths repeating an earlier one of the same agent, same cause within 16 blocks) and XP/life are per-agent means; lower is better for Deaths and Repeat, higher for XP/life. Tool era gives the highest tier reached (dmd. = diamond), the mean day of first attainment and the number of agents out of 30. Obsidian and Nether give the mean first-attainment day and the number of agents out of 30; "—" means that no agent attained it.

Table 1. Comparison with baselines on DeepSeek-V4.1-Flash.

MethodDeathsRepeatXP/lifeTool eraObsidianNether
Voyager228.9203.11.0dmd. (3.0 day, 1/30)——
Echo73.653.14.2dmd. (15.8 day, 26/30)——
Planner-only27.915.114.0dmd. (15.2 day, 16/30)——
OpenKAGE16.27.418.2dmd. (13.7 day, 24/30)22.8 day (6/30)23.8 day (4/30)

Table 2. OpenKAGE and its planner-only control across backbones.

BackboneMethodDeathsRepeatXP/lifeTool eraObsidianNether
DS-V4.1FPlanner-only27.915.114.0dmd. (15.2 day, 16/30)——
OpenKAGE16.27.418.2dmd. (13.7 day, 24/30)22.8 day (6/30)23.8 day (4/30)
GLM-5.3Planner-only33.417.48.1dmd. (15.2 day, 21/30)——
OpenKAGE28.616.924.1dmd. (13.8 day, 28/30)23.9 day (10/30)27.0 day (2/30)
hy4-previewPlanner-only47.028.88.2dmd. (16.3 day, 29/30)——
OpenKAGE15.68.332.4dmd. (18.8 day, 13/30)28.0 day (1/30)29.0 day (1/30)
Kimi-K3Planner-only64.948.77.0dmd. (21.7 day, 15/30)——
OpenKAGE42.127.06.5dmd. (19.1 day, 10/30)22.0 day (2/30)—

Table 3. Ablations on DeepSeek-V4.1-Flash. Conditions are named by what they remove: −S sharing, −R subgraph retrieval, −M the memory layer, −E the experience layer, B all graphs (the planner-only control); Raw keeps every component but shares experience records verbatim (Appendix B.5). Full is the OpenKAGE row of Table 1.

ConditionDeathsRepeatXP/lifeTool eraObsidianNether
Full16.27.418.2dmd. (13.7 day, 24/30)22.8 day (6/30)23.8 day (4/30)
Raw17.610.821.6dmd. (13.0 day, 14/30)——
−S17.48.116.1dmd. (13.3 day, 21/30)21.5 day (2/30)21.0 day (1/30)
−R15.67.413.3dmd. (13.3 day, 12/30)——
−M11.64.137.7dmd. (17.9 day, 19/30)——
−E14.57.421.7dmd. (14.0 day, 15/30)——
B27.915.114.0dmd. (15.2 day, 16/30)——

Weakness 4 / Q4. Measured costs of the main components. New Appendix B.7. Default backbone, per game night, 10 runs per cell (Tables 16–17):

Planner-onlyOpenKAGE
LLM calls (group of 3)136164
of which planning and execution136110.1
of which reflection—46.4
of which aggregation—7.5
saved text (k characters/1.7)98414
text: planning / reflection / aggregation98 / — / —155 / 221 / 38
environment steps per agent39.332.5

Reflection and aggregation account for 53 of OpenKAGE's 164 calls (32%) and 62% of its text. The cost is therefore larger in text (about 4.2×) than in calls (+21%), and a complete run takes roughly 1.6× the wall-clock time. "Text" is saved characters divided by 1.7, an estimate of tokens; billed tokens are higher because they include re-sent history. Voyager (439 calls per night) and Echo (336) both cost more per night than OpenKAGE. Other backbones (Table 16): OpenKAGE makes 12 to 59 more calls per night than its planner-only control. The paper states that the results measure progress and survival within a fixed game horizon.

Minor corrections.

  • Section 3.3 and −R: 1/3 equals 3/9, and the Flat retrieval statement is corrected. The ablation table now has 10 runs per condition; the text says that deep milestones are rare under every condition and that the ablation leaves open which component they depend on (obsidian 0/10 vs 4/10 runs for −R vs Full).
  • Table 14 "obsid.": Tool era is now read from the tool tier alone (wooden, stone, iron, diamond) and reported as tier, mean first-attainment day and count out of all agents; obsidian and the Nether have their own columns (B.2).
  • The repeated opening explanation in B.5 is removed.
  • Per-run information for Table 12 now appears in B.1 ("Sources of randomness") and in the per-run tables (launch order, seed 7).

Thank you for the review; the cost appendix and the Raw control exist because of it.

Reviewer_ao7vOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

OpenKAGE is a training-free framework for LLM agents operating in separate open-ended environments. Each agent maintains a private memory layer and a transferable experience layer to extract skills. A periodic aggregator consolidates textual experience into a centralized shared graph which influence the LLM-planning in all agents. OpenKAGE also maintains a shared registry of reusable functions, which are maintained by the originating agent and reusable by all agents. Agents retrieve textual experiences and skills based on their specific goals defined by an LLM-based planner. The paper claims that OpenKAGE enables LLM agents to choose their own objectives by leveraging global experience and then reuse existing functions to complete those objectives. Minecraft experiments show the benefits of a complete pipeline over a planner without persistent knowledge, and further results demonstrate cross-backbone reuse.

Critical strengths
  1. The architecture addresses how a group of agents can accumulate, manage, and reuse knowledge and experience. Textual experiences and transferable functions have different uses, and the paper specifies methods for consolidating experiences and functions. The separation between what agents read and what code they can execute is useful (Sections 2.2-2.4 and Appendices A.7-A.9).

  2. Table 1 reports obsidian acquisition by 6/9 and 4/9 OpenKAGE agents, versus 0/9 for each planner-only control. Nether entry occurs in 3/9 and 1/9 versus zero. The results also comment on where the baseline agents fail. Thus, the result demonstrate a benefit to deeper multi-stage progression in an open-ended Minecraft environment.

  3. The transfer experiment provides preliminary evidence that accumulated experience and skills are reusable assets and supports the need for persistence. Initializing a fresh GLM group with a complete DeepSeek group's graph and skill library shows clear improvement in reported deaths, XP and progression (Section 3.4).

  4. The paper provides clear illustrations. Figure 1 (along with Section 2.3) explicitly acknowledges that conflicting advice may be valid under different conditions, and the aggregator prompt instructs it to retain applicability conditions rather than choosing one.

  5. The paper reports the interface, prompts, graph schema, configuration and per-run results, providing details of the implementation and evaluation design choices.

Critical weaknesses

Critical weaknesses 1a. The planner-only baseline loses persistent memory, reflection skills and sharing together (Section 3.1; Appendix B.3). The comparison therefore supports the full pipeline, but does not establish that merging and reconcliing agents' textutal experience improves on simply retaining and sharing the experiences. The -S condition removes both sharing components, while -R retains consolidation and changes retrieval. Ablation is needed to demonstrate the benefits of taking conflicts into account, and whether OpenKAGE completely regresses to simpler sharing baselines as shown in Figure 1. A credible memory-equipped baseline would further contextualize the performance gains of OpenKAGE.

1b. Following this point, the baseline agents reach diamond tools, but the paper does not analyze whether later failure reflects objective selection, execution, repeated errors or loss of cross-day plans which would further explain the benefits of OpenKAGE.

2a. The authors make an effort to acknowledge the limitations in independent evidence and seed runs. However, these limitations still need to be addressed to justify validity of the results. Each main experiment has three grouped runs, while each ablated condition has one. Therefore, Table 3 results are confusing as it reports identical Nether milestone rates for Full and Flat retrieval: 3/9 and 1/3 (both 33.3%). Thus, the statement that both sharing and retrieval ablations fall below Full needs clarification. Table 12 does not document per-run seeds for model sampling or stochasiticity in retrieval. Appendix B.4 says runs differ in group identity without explaining how that determines randomization.

2b. Agents within a group run are dependent so nine agents are not necessarily nine indpendent replications of the sharing mechanism. Therefore, additional seed runs are necessary for statistical significance. Individual groups should be run with independent seeds to control for stochasticity. Additional Minecraft world seeds are also important for supporting broader robustness claims. Ideally, each method would face the same selected world seeds, with repreated groups (and seeds) on each.

  1. Generalization and efficiency remain poorly established in the current experiments. All reported runs use a single world variant (seed 7), including the transfer protocol (Table 12). These experiments are useful but they do not yet establish transfer to different world layouts and the generalizability of the acquired skills and experiences.

Game time pauses during reasoning, so a fixed game-day horizon does not match inference cost. The paper does not quantify whether the consolidation and graph traversal justify their additional calls, tokens or wall time. Additional experiments on other open-ended learning benchmarks where OpenKAGE could demonstrate broader generalization would strengthen the claims.

  1. The novelty lies primarily in consolidated group memory with textual experiences and transferable skills. Episodic/semantic graph memory already appears in "AriGraph: Learning Knowledge Graph World Models with Episodic Memory for LLM Agents, IJCAI 2025", and autonomous Minecraft exploration with executable skill accumulation and reuse appears in "Voyager: An Open-Ended Embodied Agent with Large Language Models, TMLR 2024". The paper acknowledges these connections. It's contributions could still be valuable, but a controlled comparison against these methods are needed to substantiate the methodological contributions of OpenKAGE.
Questions for authors

Clarification requests

  1. What varies between the reported group runs, and how are model-sampling and other stochastic seeds controlled? Please clarify the role of group identity and the variation across the three main-condition groups. Are the additional runs available for the ablations? With one group per ablation and one world seed through, the stability and generalizability of the conclusions remains uncertain. Please also qualify the retrieval comparison between Full and -R.

  2. Which results support sustained open-ended development beyond autonomous progression through known Minecraft milestones? Do the existing goal and execution results indicate whether shared knowledge helps agents choose subsequent objectives, execute selected objectives, or both, and where the planner-only baseline fails after reaching diamond-tier tools? This would determine whether the broader interpretation is supported or whether the claims should be narrowed.

  3. Are Full and -R matched for retrieved information/context budget, and what inference resources are used by each condition, including reflection and aggregation? Calls, tokens and wall time would help assess whether the improvement reflects the proposed organisation of knowledge and what additional the costs are, since equal game-day horizons do not imply equal inference budgets.

  4. Is the anonymous code available online? I was unfortunately unable to locate the supplementary materials for download on OpenReview.

Suggestions for future work

  1. The paper recognises that conflicting recommendations can be valid under different conditions, however, the paper does not establish how reliably LLM judgement distinguishes contextual alternatives from erroneous claims or avoids incorrect merging and deletion. An analysis would add strengthen the paper.

  2. The supplementary 60-day result is valuable but still a single-group observation. An experiment could demonstrate whether there is any advantage to maintaining more agents and whether there are diminishing returns. That being said, the computational costs of running many more LLM agents each with their own simulation, simultaneously with multiple seed runs is acknowledged.

  3. The broader research positioning is also incomplete. Parameter-level sharing is mainly represented through federated-learning-like aggregation, leaving selective decentralized reuse underrepresented. POET and OMNI would help contextualize the form of open-ended development claimed. Examples like

  • "Policy Search, Retrieval, and Composition via Task Similarity in Collaborative Agentic Systems, AAAI 2026"
  • "LLM-Powered Decentralized Generative Agents with Adaptive Hierarchical Knowledge Graph for Cooperative Planning, AAAI 2025"
  • "Selectively Sharing Experiences Improves Multi-Agent Reinforcement Learning, NeurIPS 2023"

Open-endedness work like "Paired Open-Ended Trailblazer (POET)" and "OMNI: Open-endedness via Models of human Notions of Interestingness" would help contextualize the form of open-ended development claimed in the paper.

7 comments

Thank you for a careful review. We take your points in order.

1a. The planner-only baseline removes memory, reflection, skills and sharing together. Agreed; B.3 now states that this baseline measures the combined effect of the whole stack. We added three controls, 10 runs each:

  • Raw (Table 3, B.5): everything stays as in Full (private graphs, retrieval budget, skill sharing, layer clearing); only aggregation changes. Experience records are appended to the shared graph verbatim, with merging, deduplication and conflict handling bypassed and the aggregation agent left idle. This is the condition you ask for: retaining and sharing experience without consolidating. Result vs Full: Deaths 17.6 vs 16.2, Repeat 10.8 vs 7.4, XP/life 21.6 vs 18.2; obsidian in 0/10 vs 4/10 runs; all three agents holding a diamond pickaxe in 1/10 vs 6/10 runs. At 10 runs, the survival and XP/life means of Raw and Full lie within each other's run-to-run spread. The final shared graph holds a median of 108 concepts under Full and 348 under Raw, so consolidation merges records about threefold; whether that carries into behaviour is a question for future work with more runs.
  • Voyager and Echo (Table 1): memory- and skill-equipped baselines under our protocol; deaths are far higher than under OpenKAGE (228.9 and 73.6 vs 16.2 per agent), and obsidian appears only under OpenKAGE.
  • −S is our nearest single-agent graph-memory control (private graph, retrieval and skills, no sharing). At 10 runs it stays close to Full on survival (Deaths 17.4 vs 16.2; obsidian 2/30 vs 6/30 agents).

Consolidation without conflict handling is an open variant for future work.

1b / Q2. Where does the baseline fail after diamond tools, and does shared knowledge help choosing or executing objectives? We traced the obsidian route (diamond pickaxe → iron bucket → obsidian → Nether) for every agent that holds a diamond pickaxe (OpenKAGE 24/30, planner-only 16/30).

OpenKAGEPlanner-only
proposed an obsidian goal18/249/16
median days with that goal / span of days6 / 122 / 2
iron bucket obtained13/244/16
obsidian obtained6/240/16
deaths after the pickaxe, per agent9.916.2

The planner-only agent proposes the objective (9 of 16), loses the plan across days and fails at execution. It carries no text from one day to the next (B.3), while OpenKAGE's opening prompt of about 13k characters carries yesterday's goal, outcome and memory (B.7). It also dies more while pursuing the goal.

Shared knowledge acts mainly on execution, through two channels. Textual experience records are retrieved when the plan is set (Sec. 2.1): planning with a goal triggers retrieval, so the records reach the agent once the objective is chosen and supply the how, namely preconditions, procedures and failure conditions. Executable skills act directly, since they are injected into the execution environment: on every backbone 62 to 65% of skill calls invoke a skill written by another agent (94% run without error on the default backbone), and about two thirds of the 132.5-skill library at day 30 comes from teammates (Appendix B.4). The skill libraries of the OpenKAGE runs contain skills for the obsidian route, such as filling a bucket at a water source, casting obsidian from a lava rim and building a portal frame. Objective choice is shaped indirectly, by the memory machinery that carries yesterday's goal forward (goal span 12 days vs 2); both groups propose the objective, and the difference lies in carrying it through.

Q3. Are Full and −R matched in retrieval budget? What does each condition cost? The budgets differ, and B.7 reports it. The block returned when a plan is set is 14,788 characters under Full vs 11,901 under −R (about 24% more); whole-run saved text is 414k vs 390k per night (+6%) with similar call counts (164 vs 172). −R bundles three removals (edge expansion, edge structure in results, the expand tool), so its gap reflects them jointly. Per game night on the default backbone (Tables 16–17): OpenKAGE makes 164 LLM calls vs 136 for planner-only (+21%), and 32.5 vs 39.3 environment steps per agent; saved text is about 414k vs 98k (4.2×; 62% is reflection and aggregation), estimated as saved characters/1.7, since billed tokens are higher because they include re-sent history; a complete run takes roughly 1.6× the wall-clock time. Voyager (439 calls) and Echo (336) cost more per night than OpenKAGE. The results measure progress within a fixed game horizon.

2a / Q1. What varies between runs, and how is randomness controlled? Appendix B.1 ("Sources of randomness") describes it. Fixed in every run: world seed (7), the natural spawn point of that seed, difficulty, personas, prompts, hyperparameters. Varying between runs: (i) LLM sampling (temperatures in Table 12; sampling seeds are internal to the API); (ii) the edge sampling of relational expansion, which uses an unseeded generator (anchoring is deterministic); (iii) game dynamics (mob spawns, combat, timing of asynchronous programs). Per-run seeds are therefore unavailable; the runs of a cell are repeats of one configuration, and the spread across them combines all three sources.

On Table 3: Nether 3/9 for Full and 1/3 for Flat retrieval are the same rate, so the sentence placing Flat below Full is corrected, and the table is replaced by 10 runs per condition.

2b. Dependent agents. You are right that agents within a run are dependent. All counts and comparisons use the run as the unit, and every cell now has 10 runs (the ablations too, answering your question about additional ablation runs). Default backbone, OpenKAGE vs planner-only: Deaths 16.2 vs 27.9, Repeat 7.4 vs 15.1 (all 10 OpenKAGE runs below the planner-only median in deaths), XP/life 18.2 vs 14.0; obsidian in 4/10 vs 0/10 runs and the Nether in 3/10 vs 0/10. Other backbones: mean Deaths lower on all four (GLM-5.3 28.6 vs 33.4, hy4-preview 15.6 vs 47.0, Kimi-K3 42.1 vs 64.9), and obsidian in 6/10, 1/10 and 2/10 OpenKAGE runs vs 0/10 planner-only.

3. Other world seeds. The world seed determines the terrain: the biome layout, the positions of caves, ore veins, lava lakes and water, and the natural spawn point. All three agents of a group run in separate server instances generated from the same seed, so a different seed gives the group a different world layout, with its own routes to iron, lava and water. All main experiments, including the baselines, every ablation condition and the transfer experiment, use seed 7 with the natural spawn point of that seed (Sec. 4.1, B.1). Changing the world seed or the spawn point means repeating the whole set of supporting experiments (baselines, ablations and transfer) in the new world, which calls for more time and budget than this rebuttal period allows. We therefore leave evaluation on further world layouts to future work, and the Limitations state the single-seed scope.

Other open-ended benchmarks. Our evidence covers Minecraft, and the revised Limitations state this scope.

Efficiency. See the cost numbers in our first comment to you (Tables 16–17).

Suggestions 2 and 3. The 60-day result and the group-size runs (3, 5, 8 agents) are single runs; Appendix B.6 labels them supplementary observations, endpoints at 30 days are comparable across group sizes within the observed spread, and diminishing returns are left for future work with more runs. POET, OMNI, SEPS, Nath et al. (AAAI 2026) and DAMCS (AAAI 2025) are now discussed in Related Work; the open-endedness papers concern what a single learner pursues next, the selective-sharing papers pass transitions or policy modules with tasks given in advance.

4. Novelty; controlled comparison with AriGraph and Voyager. We agree that episodic/semantic graph memory (AriGraph) and executable skill accumulation (Voyager) are prior work, and the paper cites them as such. Our contribution lies in what a group accumulates: a sharing boundary drawn by kind of knowledge (situated memories stay private; transferable text is consolidated; code is shared verbatim and stays maintained by its author), subgraph retrieval, and agents that set their own goals. Voyager: we re-implemented it under the same world, Mineflayer API (its high-level control functions removed) and horizon, three isolated agents per group (B.3). Default backbone, per agent: Deaths 228.9 vs 16.2 for OpenKAGE, diamond-tier tools 1/30 vs 24/30 agents, and obsidian and the Nether appear only under OpenKAGE. The paper describes it as a mechanism-level re-implementation under our protocol. AriGraph is outside our comparison set: it targets a single agent on text games with a given task, and an adaptation to Minecraft with executable actions would be our own re-implementation without a reference result to validate against. −S (private graph with retrieval and skills, no sharing) is the closest single-agent graph-memory control we offer.

Q4. Code. The anonymous archive, named OpenKAGE, is attached as the supplementary material of this submission: https://openreview.net/attachment?id=Neu1DIzRRT&name=supplementary_material. It contains the framework, baselines, prompts, metric and figure scripts and the regenerated figures. We apologise for the difficulty in locating it.

Suggestion 1: how reliably does consolidation separate context-dependent alternatives from errors? The paper does not measure this, and we agree it would strengthen the work. What the paper reports is the design: the aggregator tests conflicts for differing applicability conditions before correcting either side (Sec. 2.3), and it reads only experience-layer records. A measurement needs a human-labelled gold standard and a larger sample of conflicts than a 3-agent, 30-day run produces, so we leave it to future work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: What we changed

We thank the reviewers. After the initial submission we planned and re-ran a large batch of experiments, starting in October. We added method comparisons (Voyager and Echo), verified the framework on four backbones, enriched the ablations (including the new Raw condition), and sampled every cell with 10 independent runs. The PDF reports the latest numerical conclusions from these runs, and all changes relative to the initial submission are marked in blue text. The larger sample refined some earlier claims, and the PDF states the refined versions.

What changed in the PDF

  • Sample size. Every cell now has 10 independent group runs of 3 agents (the submission had 3 groups per main cell and 1 per ablation). All seven ablation conditions have 10 runs. Because the agents of a run share one graph, the run is the sampling unit (Appendix B.2). Per-run values are in Table 13, Table 15 and Fig. 9. These replies report cell means and run counts; the per-run tables allow recomputation.
  • Controls. Voyager and Echo are re-implemented under the same worlds, low-level Mineflayer API and horizon (Table 1, B.3). A new Raw condition shares experience records verbatim without consolidation (Table 3, B.5).
  • Cost. New Appendix B.7 (Tables 16–17): LLM calls, steps, text volume per component, wall-clock.
  • Randomness. Appendix B.1 lists what is fixed and what varies between runs.
  • Raw data. The raw run data behind the conclusions of the paper is available at https://github.com/zhuzhuxia5678/OpenKAGE_results.
  • Related work. Added POET, OMNI, G-Memory, LLMA-Mem (arXiv 2604.03295), SEPS, Nath et al., DAMCS; two rows added to Table 11. What the larger sample shows
  • Survival is the robust effect. Default backbone, OpenKAGE vs planner-only: Deaths 16.2 vs 27.9, repeated deaths 7.4 vs 15.1; all 10 OpenKAGE runs lie below the planner-only median in deaths. Deaths are lower on all four backbones.
  • Exploration is where the gap is clearest. Across the four backbones, obsidian and the Nether appear only under OpenKAGE: 0 of 120 planner-only agents, 0 of 30 Voyager agents and 0 of 30 Echo agents reach either, while OpenKAGE obtains obsidian on all four backbones (6, 10, 1 and 2 of 30 agents; in 4, 6, 1 and 2 of 10 runs) and enters the Nether on three.
  • The ablations reflect the components' contributions to a degree. Every ablated condition keeps deaths well below the planner-only control (11.6 to 17.6 vs 27.9 per agent), and on the deepest milestones each falls below Full: obsidian in 0 to 2 of 30 agents against 6, the Nether in 0 to 1 against 4. −M has fewer deaths (11.6 vs 16.2) and higher XP/life (37.7 vs 18.2) than Full, with 0 obsidian and 0 Nether agents. Apportioning the effect to each component needs more runs: Full itself reaches these milestones in 4 of 10 runs, the survival means of −S, −R, −E and Raw lie within Full's run-to-run spread, and the conditions overlap by design (−S and −E both leave the shared graph empty; −R bundles three removals). Ablation claims that rested on single runs are removed from the PDF.
  • Scope. The claims cover Minecraft and 30 game days (60 in one supplementary run); within this setting, we study a framework that drives lifelong learning and collective evolution of agents.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.