MemPlan: Memory-Conditioned PDDL Planning for Partially Observable Text Environments

Tengwei Song; Robert Hoehndorf

cs.CLOct 21, 2026

MemPlan: Memory-Conditioned PDDL Planning for Partially Observable Text Environments

Authors: Tengwei Song, Robert Hoehndorf

Organizations: King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia · Åbo Akademi University, Finland · ELLIS Institute Finland

Abstract

LLM agents are flexible in language interaction, but using them as long-horizon planners is costly and prone to hallucinated or invalid actions. This paper targets a different role for LLMs: constrained bridges between language and symbolic planning. We propose MemPlan, a memory-conditioned PDDL planning framework for partially observable interactive text environments. MemPlan keeps the PDDL domain fixed and reconstructs the PDDL problem at each step from parsed observations, candidate hypotheses, and interaction memory. Positive graph memory assigns relevance-based action costs to candidate hypotheses for missing objects, locations, or conditions, guiding the planner toward more plausible checks. Negative memory records execution-refuted assumptions and compiles them into exclusion constraints, preventing repeated attempts at candidates already contradicted by feedback. A unified schema-conditioned language--symbol interface, trained offline with supervised fine-tuning, connects text environments to the planning loop through observation-to-fact parsing and symbolic-action-to-command grounding. Experiments on TextWorld, ALFWorld, and Robotouille show that MemPlan improves task completion and step efficiency over language-only and PDDL-aware baselines, while substantially reducing online token cost compared with LLM-planning baselines such as ReAct. The same fine-tuned interface is also reused across TextWorld task types without task-specific fine-tuning. Our code and data are available at: https://github.com/tewiSong/MemPlan.

Figures & tables

Figure 1: MemPlan online planning loop. Given the task goal, current observation, and admissible commands, the language–symbol interface extracts symbolic facts and grounds planner actions. MemPlan builds a PDDL problem from the fixed domain, candidate hypotheses, positive memory, and negative memory. Environment feedback updates memory and triggers replanning.

Model	clean	cool	examine	heat	put	puttwo	overall $\uparrow$	length $\downarrow$	tokens $\downarrow$
LLM-DP	0.94	1.00	1.00	0.87	1.00	0.94	0.96	13.16	633k
GPT-4o	0.13	0.19	0.39	0.09	0.38	0.12	0.21	127.18	6.79M
GPT-4o + Domain	0.90	0.90	0.83	0.91	0.83	0.71	0.86	34.69	13.09M
ReAct (GPT-4o)	0.74	0.90	0.78	0.87	0.88	0.82	0.83	38.18	5.07M
ReAct (LLM-DP)	0.35	0.90	0.33	0.65	0.71	0.29	0.54	18.69	9.16M
MemPlan	1.00	1.00	1.00	1.00	1.00	1.00	1.00	6.91	906k

Table 1: ALFWorld task completion, episode length, and online LLM token usage.

Method	Success (%) $\uparrow$	Steps $\downarrow$	Tokens $\downarrow$
ReAct (GPT-5.1)	98.5	10.92	2.74M
PDDLEGO+-Native	94.8	14.32	2.31M
PDDLEGO+-Shared	92.5	15.47	1.82M
LLM-DP	96.0	13.16	0.633M
MemPlan	100	6.91	0.906M

Table 2: ALFWorld comparison with frontier LLM and PDDL-aware baselines (134 episodes).

Figure 2: TextWorld normalized scores against memory-augmented LLM agents.

Model	Synchronous (%) $\uparrow$	Asynchronous (%) $\uparrow$
ReAct (GPT-4o)	47.0	11.0
ReAct (GPT-5.1)	63.6	24.7
LLM-DP	57.9	18.5
ReAct (GPT-4o-mini)	11.0	0.0
ReAct (Gemini-1.5-Flash)	0.0	0.0
ReAct (Claude-3-Haiku)	2.0	0.0

Table 3: Robotouille task success in synchronous and asynchronous settings. Values are macro-averaged over task configurations after averaging over seeds within each configuration.

Model	Norm. score $\uparrow$	Steps $\downarrow$	Rep. checks / run $\downarrow$	Term. dead ends / run $\downarrow$	Invalid acts. / run $\downarrow$
MemPlan	1.00	31.0	0.0	0.0	0.0
Uniform-cost hypotheses	1.00	38.0	1.0	0.0	0.0
w/o negative memory	0.64	60.0	12.0	0.0	1.0
w/o placeholders	0.27	–	0.0	1.0	0.0
w/o observation parser	0.55	60.0	3.0	0.0	4.0
w/o action grounder	0.45	60.0	1.0	0.0	9.0

Table 4: TextWorld ablations. Values are means over 21 runs. Step count is omitted for terminal dead-end variants, where shorter length does not imply higher efficiency.

Model	Steps $\downarrow$	Planner calls $\downarrow$	LLM calls $\downarrow$	Tokens / episode $\downarrow$	Invalid action rate $\downarrow$
ReAct (GPT-4o)	38.18	–	39.78	37.9k	0.03
LLM-DP	13.16	13.16	13.16	4.7k	0.01
MemPlan	6.91	6.91	16.87	6.9k	0.00

Table 5: ALFWorld online efficiency breakdown. Invalid-action rate is the fraction of executed commands rejected by the environment.

Panel of Figure 3 — Figure 3: Case study on a TextWorld Cooking Hard task with an unresolved grilling device. MemPlan introduces a planner-side placeholder unknown_grill_1 , uses feedback to avoid treating an unrealized symbolic action as executed, replans, and grounds the resolved device bbq_0 to the executable command cook red tuna with BBQ .

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Environment	Evaluation unit	Episode budget
TextWorld	7 task instances	task-dependent limits
ALFWorld	134 episodes	150 steps
Robotouille	151 completed multi-seed runs	derived from each task’s optimal-step budget

Appendix

Table 6: Evaluation runs and episode budgets.

Group in Figure 2	Instance identifiers	Role in transfer evaluation
Treasure Hunt Hard	hunt, hunt_hard	fine-tuning task family
Cooking Hard	cook, cook_hard, cook_hardest, cook_rl_baseline	fine-tuning task family
Cleaning	clean_3x3	held-out task family

Appendix

Table 7: TextWorld evaluation grouping.

Environment	Resampling unit	Main metric	Main-text estimate	Appendix statistics
ALFWorld	episode	success, length, tokens	yes	95% CI, paired differences
TextWorld	task instance	normalized score	yes	per-instance scores, group averages
Robotouille	completed run	success	yes	95% CI by sync/async setting
Ablations	episode or run	success, length, diagnostics	yes	95% CI, paired differences

Appendix

Table 8: Statistical reporting summary.

Model	Action-generation tokens	Interface tokens	Planner tokens
GPT-4o	yes	no	no
GPT-4o + Domain	yes	no	no
ReAct	yes	no	no
LLM-DP	yes	no	no
MemPlan	no	yes	no

Appendix

Table 9: Online token accounting. MemPlan uses language-model calls only for parsing and grounding.

Model	Steps / episode	LLM calls / episode	Tokens / episode	95% CI for tokens / episode
ReAct (GPT-4o)	38.18	39.78	37.9k	[34.8k, 41.2k]
LLM-DP	13.16	13.16	4.7k	[4.2k, 5.3k]
MemPlan	6.91	16.87	6.9k	[6.4k, 7.5k]

Appendix

Table 10: Per-episode online efficiency statistics.

Training task format	Number of examples	Label source
Observation-to-fact parsing	17,000	simulator annotations
Symbolic-action-to-command grounding	9,600	planner-action / oracle-trajectory / admissible-command alignment

Appendix

Table 11: Training data for the unified Qwen2.5-7B language–symbol interface.

Fine-tuning task families	Evaluation task groups in Figure 2	Interface update during evaluation
cook, hunt	Cooking Hard, Treasure Hunt Hard, Cleaning	none

Appendix

Table 12: TextWorld interface-transfer setting. Cooking Hard and Treasure Hunt Hard correspond to the fine-tuning task families, while Cleaning evaluates cross-task use of the frozen interface.

Environment / split	Parsing precision	Parsing recall	Parsing F1	Grounding top-1	Schema rejection	Env. rejection
ALFWorld validation	0.94	0.92	0.93	0.96	0.03	0.01
TextWorld validation	0.96	0.94	0.95	0.97	0.02	0.01
Robotouille validation	0.91	0.89	0.90	0.94	0.04	0.02

Appendix

Table 13: Validation metrics for the schema-conditioned language–symbol interface.

Split	Parser omission	Parser hallucination	Entity mismatch	Predicate mismatch	Grounding mismatch	No safe grounding
ALFWorld	28	14	18	10	20	10
TextWorld	24	12	16	9	28	11
Robotouille	30	16	20	12	14	8

Appendix

Table 14: Interface error breakdown.

Item	Value
Base model	Qwen2.5-7B-Instruct
Fine-tuning method	LoRA, rank 16, alpha 32, dropout 0.05
Training epochs	2
Learning rate	$5\times 10^{-5}$
Batch size	4 per device, gradient accumulation 4
Maximum input length	131,072 tokens tokenizer limit

Appendix

Table 15: Training hyperparameters for the unified Qwen2.5-7B interface model.

Model	Goal / observation	Interaction history	Admissible commands	Domain information	Symbolic planner
GPT-4o	yes	yes	benchmark-dependent	no	no
GPT-4o + Domain	yes	yes	benchmark-dependent	natural-language domain description	no
ReAct	yes	recent history	benchmark-dependent	no	no
LLM-DP	yes	yes	benchmark-dependent	PDDL-aware belief planning	yes
PDDLEGO / PDDLEGO+	yes	yes	benchmark-dependent	iterative PDDL construction/refinement	yes
MemPlan	yes	symbolic memory	yes when exposed	fixed PDDL domain and predicate schema	yes

Appendix

Table 16: Information available to each method at test time. “Benchmark-dependent” means that admissible-command access is used only when the corresponding benchmark exposes it and is applied consistently across compared methods within that benchmark.

Model / component	Model version	Max output tokens	History window
GPT-4o	gpt-4o	512	full episode history
GPT-4o + Domain	gpt-4o	512	full episode history + domain description
ReAct (GPT-4o)	gpt-4o	512	full episode history
ReAct (LLM-DP)	gpt-4o	512	belief-state history
MemPlan parser	Qwen2.5-7B fine-tuned	1024	current observation
MemPlan grounder	Qwen2.5-7B fine-tuned	1024	current observation + admissible commands

Appendix

Table 17: Prompt and decoding configuration.

LLM Agent Planning

Large language model (LLM) agent planning studies how agents decompose goals, choose actions and tools, and adapt plans from feedback.

See all 225 papers

Agentic LLM Planning via Step-Wise PDDL Simulation: An Empirical Characterisation
Mar 6, 2026Kai Göbel, Pierrick Lorang, Patrik Zips +1LLM Planning LLM Agent Planning
AIT Austrian Institute of Technology GmbH, Vienna, Austria
PDDLCoder: Agentic PDDL Generation for LLM-Assisted Symbolic Planning
Aug 17, 2026Veit Laule, Jiangtao Shuai, Manfred Hauswirth +1LLM Planning Symbolic Planning
Technical University of Berlin Berlin, Germany · Fraunhofer FOKUS & Technical University of Berlin Berlin, Germany
Planning Commitment of LLM under Partial Observability
Oct 15, 2026Cassie Huang, Zhanwei Cao, Yeawon Lee +6LLM Planning LLM Agent Safety

Mar 6, 2026J/K move · Enter open · S save

Agentic LLM Planning via Step-Wise PDDL Simulation: An Empirical Characterisation

Kai Göbel, Pierrick Lorang, Patrik Zips, Tobias Glück

AIT Austrian Institute of Technology GmbH, Vienna, Austria

LLM Planning LLM Agent Planning Large Language Model-Based Robot Planning Symbolic Planning

Open paper AskPDF arXiv

Task planning, the problem of sequencing actions to reach a goal from an initial state, is a core capability requirement for autonomous robotic systems. Whether large language models (LLMs) can serve as viable planners alongside classical symbolic methods remains an open question. We present PyPDDLEngine, an open-source Planning Domain Definition Language (PDDL) simulation engine that exposes planning operations as LLM tool calls through a Model Context Protocol (MCP) interface. Rather than committing to a complete action sequence upfront, the LLM acts as an interactive search policy that selects one action at a time, observes each resulting state, and can reset and retry. We evaluate four approaches on 102 International Planning Competition (IPC) Blocksworld instances under a uniform 180-second budget: Fast Downward lama-first and seq-sat-lama-2011 as classical baselines, direct LLM planning (Claude Haiku 4.5), and agentic LLM planning via PyPDDLEngine. Fast Downward achieves 85.3% success. The direct and agentic LLM approaches achieve 63.7% and 66.7%, respectively, a consistent but modest three-percentage-point advantage for the agentic approach at

5.7\times

higher token cost per solution. Across most co-solved difficulty blocks, both LLM approaches produce shorter plans than seq-sat-lama-2011 despite its iterative quality improvement, a result consistent with training-data recall rather than generalisable planning. These results suggest that agentic gains depend on the nature of environmental feedback. Coding agents benefit from externally grounded signals such as compiler errors and test failures, whereas PDDL step feedback is self-assessed, leaving the agent to evaluate its own progress without external verification.

Abstract

LLM Agent Planning

Explore similar work

Agentic LLM Planning via Step-Wise PDDL Simulation: An Empirical Characterisation

PDDLCoder: Agentic PDDL Generation for LLM-Assisted Symbolic Planning

Planning Commitment of LLM under Partial Observability