AuthorsHongxin Li, Zeyu Wang, Xu Yang et al.
TL;DR
MemoNav uses a working-memory-style combination of STM, LTM, and a selective forgetting module to reach 74.7% SR on Gibson 1-goal, +4.7 points over VGM.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing ImageNav methods use all historical observations for decision-making without considering the goal-relevant fraction, leading to inefficient exploration.
In image-goal navigation on Gibson and Matterport3D, this means agents traverse many redundant nodes, degrading navigation efficiency and lowering success rate on long multi-goal episodes.
HOW IT WORKS
MemoNav’s core mechanism combines Short-term memory, a Selective forgetting module, Long-term memory, Working memory generation, and Transformer decoders over a topological map.
You can think of MemoNav like a brain with a notepad (STM), a long-term summary (LTM), and an attention filter that erases unhelpful notes while keeping a compact, task-focused scratchpad (WM).
This working-memory-style design lets MemoNav focus on goal-relevant nodes and scene-level structure instead of a flat context window over all past observations.
DIAGRAM
This diagram shows how MemoNav processes images and updates STM, LTM, and WM at each time step before generating an action.
DIAGRAM
This diagram shows how MemoNav is trained on Gibson 1-goal, then evaluated with ablations on Gibson and Matterport3D multi-goal tasks.
PROCESS
STM generation
MemoNav uses the memory update module to encode the current panoramic RGBD image and store landmark node features as Short-term memory on the topological map.
Selective forgetting
MemoNav’s selective forgetting module ranks STM nodes by attention scores from Dgoal and temporarily removes nodes below threshold p from subsequent decision-making.
LTM generation
MemoNav maintains a trainable global node as Long-term memory that connects to all STM nodes and progressively aggregates their features at each time step.
Working memory generation and action
MemoNav applies GATv2-based working memory generation over retained STM and LTM, then two Transformer decoders and a policy network convert WM into navigation actions.
KEY CONTRIBUTIONS
MemoNav working memory model
MemoNav introduces three scene representations—Short-term memory, Long-term memory, and Working memory generation—to improve image-goal navigation across Gibson and Matterport3D multi-goal tasks.
Selective forgetting module
MemoNav’s selective forgetting module uses attention scores from Dgoal to retain only informative STM, reducing redundancy and enabling higher SR and PR with fewer active nodes.
Global LTM node with GATv2 WM
MemoNav adds a global LTM node and a GATv2-based working memory generator, yielding up to +8.5 PR points over VGM on Gibson 3-goal tasks and +7.4 on 4-goal tasks.
RESULTS
SR (Gibson 1 goal)
74.7%
+4.7 over VGM
SPL (Gibson 1 goal)
57.9%
+2.5 over VGM
PR (Gibson 4 goal)
28.9%
+7.4 over VGM
PR (Matterport3D 3 goal)
13.6%
+1.8 over VGM
On Gibson and Matterport3D ImageNav benchmarks, MemoNav is evaluated on 1-goal and multi-goal tasks, showing consistent SR and PR gains over VGM and TSGM. These numbers demonstrate that MemoNav’s working-memory-style STM, LTM, and selective forgetting pipeline yields more successful and efficient long-horizon navigation.
BENCHMARK
On Gibson and Matterport3D ImageNav benchmarks, MemoNav is evaluated on 1-goal and multi-goal tasks, showing consistent SR and PR gains over VGM and TSGM. These numbers demonstrate that MemoNav’s working-memory-style STM, LTM, and selective forgetting pipeline yields more successful and efficient long-horizon navigation.
BENCHMARK
Success Rate (SR) on Gibson 1-goal hard episodes.
BENCHMARK
Progress (PR) on Gibson 2-goal tasks for VGM and MemoNav component variants.
KEY INSIGHT
MemoNav maintains high PR on 3-goal tasks even when retaining only 20% of STM nodes, yet still reaches strong success rates.
This is surprising because one might expect aggressive forgetting to cripple navigation, but MemoNav’s attention-based selection shows most nodes are unnecessary for efficient multi-goal planning.
WHY IT MATTERS
MemoNav shows that navigation agents can use working-memory-style STM, LTM, and WM to plan efficient paths with compact, goal-focused scene memory.
Builders can now design embodied agents that scale to long multi-goal missions in complex 3D environments without exploding memory or exhaustive re-exploration.
Nikhil Verma
· 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 → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 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 MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 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.
Checking…