Active Memory Navigation

Updated 10 July 2026

Active Memory Navigation is a framework that integrates memory construction, retrieval, and replanning as part of an agent's continuous decision-making process.
It employs diverse memory structures, including non-parametric graphs, episodic buffers, and hierarchical pyramids, to enhance localization and adaptive planning.
Practical implementations show significant improvements in long-horizon navigation tasks by reducing deadlocks and boosting success metrics in zero-shot and lifelong scenarios.

Active memory navigation denotes a class of methods in which an agent treats memory access not as a one-shot lookup tool but as an integral part of its reasoning loop, interleaving memory construction, retrieval, verification, replanning, and action. In the literature, the term spans embodied navigation in previously unseen environments, structured traversal over conversational memory, and hierarchical exploration of document collections. Its concrete realizations include non-parametric topological graphs, episodic semantic buffer queues, short-term/long-term/working-memory pipelines, persistent 3D Gaussian maps, hierarchical long-short memory systems, and multi-granularity memory pyramids (Li et al., 27 May 2026, Savinov et al., 2018, Wu et al., 25 Mar 2026, Li et al., 2024, He et al., 24 Jun 2025, Xu et al., 7 Jul 2026).

1. Problem setting and research lineage

In embodied AI, the central motivation is that navigation tasks are long-horizon, partially observable, and often open-world. Zero-shot object navigation requires agents to locate unseen target objects in unfamiliar environments without prior maps or task-specific training, while image-goal navigation requires an agent to navigate to a goal indicated by an image in unfamiliar environments. Lifelong embodied navigation further requires agents to accumulate, retain, and exploit spatial-semantic experience across tasks. Across these settings, flat memory use is repeatedly identified as a bottleneck: recent VLM-based systems still suffer from spatial hallucinations, local exploration deadlocks, and a disconnect between high-level semantic intent and low-level control, while earlier image-goal methods often used all observations in the memory for generating navigation actions without considering which fraction of this memory is informative (Wu et al., 25 Mar 2026, Li et al., 2022, Ren et al., 25 Dec 2025).

A foundational line in embodied navigation is the semi-parametric topological approach. "Semi-parametric Topological Memory for Navigation" stores locations as nodes in a non-parametric graph and uses a parametric retrieval network for localization and loop closure. The graph stores no metric information, only connectivity of locations corresponding to the nodes. Given only 5 minutes of footage of a previously unseen maze, the SPTM-based agent builds a topological map and uses it to navigate toward goals; on the reported test environments, the average success rate is higher than the best-performing baseline by a factor of three, and the detailed evaluation reports 100% success in all test mazes (Savinov et al., 2018).

A second line replaces explicit planning modules with learned policies conditioned on episodic memory. "Memory-Augmented Reinforcement Learning for Image-Goal Navigation" uses a self-supervised state-embedding network, a write-only episodic buffer with capacity $C=20$ , and transformer-style attention over the buffer, achieving overall $0.69/0.56$ success/SPL on Gibson from RGB alone (Mezghani et al., 2021). "MemoNav: Selecting Informative Memories for Visual Navigation" and the later "MemoNav: Working Memory Model for Visual Navigation" introduce a three-store design in which short-term memory, long-term memory, and working memory are explicitly separated, with a forgetting module that retains the informative fraction of short-term memory (Li et al., 2022, Li et al., 2024).

More recent systems diversify the form of memory rather than abandoning it. DGMem incrementally builds a Dynamic Graph Memory from on-board observations and uses the same memory for planning, intrinsic reward generation, and imitation data, without any labels (Cai et al., 2023). Memory Proxy Maps replace graph and odometry with a learned latent occupancy-style proxy map and a feudal hierarchy, explicitly described as a no-RL, no-graph, no-odometry, and no metric map approach (2411.09893). StateLinFormer shifts attention from memory structure to memory persistence, training a linear-attention navigation model with stateful memory carried across consecutive training segments to approximate learning on infinitely long sequences (Chen et al., 24 Mar 2026). These trajectories suggest that active memory navigation is less a single architecture than a recurring systems principle.

2. Memory representations and data structures

The literature instantiates active memory navigation through markedly different memory substrates. The following representations recur across the cited systems.

Framework	Memory representation	Role
SPTM	non-parametric graph $G=(V,E)$ + retrieval network	localization, loop closure, shortest-path planning
ReMemNav	Episodic Semantic Buffer Queue $\mathcal{M}$ + panoramic semantic priors	rethinking, deadlock prevention, target verification
MemoNav	STM, LTM, WM	goal-relevant scene selection in a topological map
3DGSNav	3D Gaussian Splatting persistent memory	free-viewpoint rendering and frontier-aware reasoning
Mem4Nav	sparse octree + semantic topology graph + STM/LTM	fine-grained voxel indexing and landmark connectivity
NapMem	linked multi-granularity memory pyramid	memory as a structured action space

In SPTM, nodes correspond to observations from an exploration video, and edges are added either for temporal adjacency or for visually similar loop closures. Planning consists of localizing the current observation and the goal in the graph, running Dijkstra, and selecting a reachable waypoint on the shortest path (Savinov et al., 2018). ReMemNav instead uses a sliding-window list of key-value pairs,

$\mathcal{M}_t=\{(p_i,d_i)\},$

where $p_i\in\mathbb{R}^2$ is the $(x,y)$ world coordinate and $d_i$ is the panoramic spatial description generated by the VLM; its capacity is $K=10$ , and the queue is first-in, first-out. At each time step, panoramic semantic priors are built from six discrete yaw angles $\theta_i\in\{30^\circ,90^\circ,150^\circ,210^\circ,270^\circ,330^\circ\}$ using the Recognize Anything Model, and the resulting dictionary is fed jointly with the stitched panorama to the VLM (Wu et al., 25 Mar 2026).

MemoNav’s representation is explicitly cognitive in nomenclature. Short-term memory is the set of node features in the topological map; long-term memory is a single global memory node that progressively aggregates short-term memory via GATv2; working memory is the union of the retained short-term memory and the updated long-term memory (Li et al., 2024). Mem4Nav generalizes the split-memory pattern to urban VLN: long-term memory compresses and retains historical observations at octree and graph nodes through reversible Transformer tokens, while short-term memory caches recent multimodal entries in relative coordinates for local planning (He et al., 24 Jun 2025).

Outside embodied navigation, memory becomes a navigable graph or pyramid. MemCog organizes user knowledge as a Navigable Memory Store with associative link graphs and a three-level hierarchy of dimension, page, and section (Li et al., 27 May 2026). NapMem organizes raw conversations, typed memory records, topic tracks, and a user profile into a directed acyclic graph with provenance edges, exposing each layer through memory tools rather than a single retrieval interface (Xu et al., 7 Jul 2026). NaviRAG similarly restructures documents into a hierarchical memory $0.69/0.56$0, with summaries at internal nodes and raw segments at the leaves, so that retrieval can move between coarse and fine granularity (Dai et al., 14 Apr 2026). This range of representations indicates that active memory navigation is compatible with graph, queue, token, latent-map, and hierarchical-file abstractions.

3. Closed-loop memory operations

A defining property of active memory navigation is that memory is queried, updated, and revised inside the control loop. In SPTM, the agent repeatedly localizes the current observation, recomputes a shortest path, selects a waypoint not too close and not too far, executes a locomotion action, and repeats until the goal is reached. The original paper explicitly characterizes this as an agent that can “actively” plan by re-querying its memory at each step, re-planning a shortest path, selecting a reachable waypoint, and invoking low-level control (Savinov et al., 2018).

ReMemNav makes this loop explicit through adaptive dual-modal rethinking. Before rethinking, the VLM outputs a JSON-structured response

$0.69/0.56$1

where $0.69/0.56$2 is one of the six angles, $0.69/0.56$3 is a True/False goal flag, $0.69/0.56$4 is natural-language reasoning, and $0.69/0.56$5 is a panoramic spatial description. Mode A, Decision Evaluation, is used when $0.69/0.56$6. It projects past memory nodes into the current camera view, triggers when projected nodes fall inside the collision-free navigation mask and satisfy the angular threshold, retrieves the furthest hit node, and re-prompts the VLM with the stored description to avoid seen areas. Mode B, Target Verification, is used when $0.69/0.56$7: the system re-asks the VLM to verify the single-view image at angle $0.69/0.56$8, and if verification fails it sets $0.69/0.56$9 and falls back to Mode A. The gating is summarized as

$G=(V,E)$ 0

Low-level action selection is then delegated to a second VLM-mediated choice over depth-mask rays (Wu et al., 25 Mar 2026).

A related loop appears in memory-centric conversational and retrieval systems. MemCog formalizes active memory navigation as interleaved Think–Act–Observe cycles over memory, where the agent chooses among actions such as list_dimensions, browse_dimension, read_page, follow_link, or answer, updating its internal state after each tool output (Li et al., 27 May 2026). NapMem casts the process as an episodic Markov decision process in which actions are memory-tool invocations at granularity levels $G=(V,E)$ 1, $G=(V,E)$ 2, $G=(V,E)$ 3, and $G=(V,E)$ 4, or termination with Answer, and trains the policy with Group Relative Policy Optimization (Xu et al., 7 Jul 2026). NaviRAG uses a related but document-centric loop: it maintains candidate nodes $G=(V,E)$ 5, selects top-scoring nodes under a relevance score $G=(V,E)$ 6, and decides for each node whether to absorb its summary or expand to its children (Dai et al., 14 Apr 2026).

Active memory can also be embedded directly into recurrent state rather than surfaced through explicit tool calls. StateLinFormer maintains a memory matrix $G=(V,E)$ 7 with linear-attention update

$G=(V,E)$ 8

and its stateful training regime carries memory across consecutive training segments rather than resetting it at each batch boundary. The reported interpretation is that this exposes the model to memory states approximating the recurrent dynamics under an infinitely long sequence (Chen et al., 24 Mar 2026). This suggests that active memory navigation includes both explicit traversal over external memory and implicit persistence in recurrent latent state.

In current embodied work, ReMemNav is a canonical training-free zero-shot formulation. The system integrates panoramic semantic priors, an episodic semantic buffer queue, an adaptive dual-modal rethinking mechanism, and low-level action selection based on feasible rays from depth masks. The queue capacity sweep using Qwen3-VL-4B shows an inverted-U with respect to $G=(V,E)$ 9, peaking at $\mathcal{M}$ 0: too small provides no context to break loops, while too large introduces noise from distant context. On the main zero-shot evaluation, ReMemNav reports MP3D $\mathcal{M}$ 1, $\mathcal{M}$ 2; HM3D v0.1 $\mathcal{M}$ 3, $\mathcal{M}$ 4; and HM3D v0.2 $\mathcal{M}$ 5, $\mathcal{M}$ 6, corresponding to absolute gains of $\mathcal{M}$ 7, $\mathcal{M}$ 8, and $\mathcal{M}$ 9, respectively (Wu et al., 25 Mar 2026).

Other object-navigation systems instantiate the same principle with different memory substrates. 3DGSNav uses a persistent 3D Gaussian Splatting memory, active perception through panoramic opacity clustering, frontier-aware first-person rendering, and VLM-driven re-verification. On HM3Dv1, HM3Dv2, and MP3D, it reports $\mathcal{M}_t=\{(p_i,d_i)\},$ 0 of $\mathcal{M}_t=\{(p_i,d_i)\},$ 1, $\mathcal{M}_t=\{(p_i,d_i)\},$ 2, and $\mathcal{M}_t=\{(p_i,d_i)\},$ 3, with an average improvement of $\mathcal{M}_t=\{(p_i,d_i)\},$ 4 SR and $\mathcal{M}_t=\{(p_i,d_i)\},$ 5 SPL over prior SOTA baselines; the real-robot evaluation reports overall success $\mathcal{M}_t=\{(p_i,d_i)\},$ 6 (Zheng et al., 12 Feb 2026). MAG-Nav combines perspective-based active grounding with historical memory backtracking. On GOAT-Bench in HM3D, the reported full system reaches Success $\mathcal{M}_t=\{(p_i,d_i)\},$ 7, SPL $\mathcal{M}_t=\{(p_i,d_i)\},$ 8, and DTG $\mathcal{M}_t=\{(p_i,d_i)\},$ 9, improving over VLN-Game with GPT-4o at $p_i\in\mathbb{R}^2$ 0 (Zhang et al., 7 Aug 2025).

In image-goal and multi-goal navigation, MemoNav frames active memory navigation as selective retention. The 2022 MemoNav reports that the model outperforms SoTA methods by a large margin with a smaller fraction of navigation history and is less likely to be trapped in a deadlock (Li et al., 2022). The 2024 MemoNav reports Gibson 1-goal $p_i\in\mathbb{R}^2$ 1, $p_i\in\mathbb{R}^2$ 2 against VGM’s $p_i\in\mathbb{R}^2$ 3, with consistent gains across multi-goal Gibson and Matterport3D settings, and attributes these gains to the synergy among STM, LTM, and WM (Li et al., 2024). Memory Proxy Maps arrive at similar goals through a different route: the memory is a compact occupancy-style map in a learned 2D proxy space, and the overall feudal agent reports Straight Success $p_i\in\mathbb{R}^2$ 4, SPL $p_i\in\mathbb{R}^2$ 5, and Curved Success $p_i\in\mathbb{R}^2$ 6, SPL $p_i\in\mathbb{R}^2$ 7, exceeding the compared baselines while using no RL, no graph, no odometry, and no metric map (2411.09893).

Long-horizon and lifelong variants emphasize persistence or compression. StateLinFormer improves MAZE success from $p_i\in\mathbb{R}^2$ 8 to $p_i\in\mathbb{R}^2$ 9 and ProcTHOR success from $(x,y)$ 0 to $(x,y)$ 1 when moving from stateless to stateful training on the same linear-attention architecture (Chen et al., 24 Mar 2026). AstraNav-Memory compresses each image to about 30 tokens under a representative $(x,y)$ 2 compression setting and reports GOAT-Bench Val-Unseen $(x,y)$ 3, $(x,y)$ 4 and HM3D-OVON Val-Unseen $(x,y)$ 5, $(x,y)$ 6, while showing that moderate compression provides the best balance between efficiency and accuracy (Ren et al., 25 Dec 2025). At the biological extreme, the spiking neuron-astrocyte network combines STDP with astrocytic short-term suppression and reports up to sixfold median path-length reduction together with large success gains under extreme partial observability (Tsybina et al., 16 Apr 2026).

The phrase active memory navigation has been generalized beyond robot motion. MemCog defines it as the process by which a conversational agent treats memory access not as a one-shot lookup tool but as an integral part of its reasoning loop. Its system exposes a Cross-Dimensional Navigation Interface over a Navigable Memory Store and adds a Proactive Reasoning Protocol so that memory exploration can begin even without an explicit user question. Quantitatively, the reported full system reaches $(x,y)$ 7 on LoCoMo, $(x,y)$ 8 on LongMemEval, and $(x,y)$ 9 with LLM-judged precision $d_i$ 0 on ProactiveMemBench (Li et al., 27 May 2026).

NapMem pushes the same idea into a reinforcement-learning setting. It organizes user history into a linked multi-granularity memory pyramid connecting raw conversations, typed memory records, topic tracks, and a user profile, and trains the agent to select memory according to the query and intermediate evidence. On memory-intensive tasks, the reported NapMem-9B with RL reaches an average score of $d_i$ 1, compared with $d_i$ 2 for the variant without RL and $d_i$ 3 for the variant without navigation. On non-memory tasks, BFCL-v3 task accuracy remains $d_i$ 4, while unnecessary memory calls on GPQA-Diamond are reported as $d_i$ 5 (Xu et al., 7 Jul 2026). This suggests that active memory navigation is being treated not only as a storage problem but as a policy-learning problem.

NaviRAG extends the paradigm to retrieval-augmented generation. It structures documents into a hierarchy of topics, subtopics, and raw segments, then lets an LLM agent actively navigate this hierarchy, iteratively identifying information gaps and retrieving the most appropriate granularity. On NarrativeQA, the reported F1 improves from vanilla RAG’s $d_i$ 6 to $d_i$ 7, and on LongBench-v2 the accuracy increases from $d_i$ 8 to $d_i$ 9. The ablation without the navigation module drops NarrativeQA F1 from $K=10$ 0 to $K=10$ 1, while removing the hierarchical KB drops it to $K=10$ 2 (Dai et al., 14 Apr 2026).

Web navigation offers a complementary memory design. ReAP stores natural-language reflections from successful and failed trajectories, retrieves the top- $K=10$ 3 most similar reflections by cosine similarity, and prepends them to the LLM’s prompt. On unseen WebArena tasks, the baseline success rate of $K=10$ 4 rises to $K=10$ 5 with Summary or Reflection, and on previously failed held-out tasks the success rate rises from $K=10$ 6 to $K=10$ 7; steps are reduced by $K=10$ 8 overall with Reflection (Azam et al., 2 Jun 2025). Human memory support systems have also adopted analogous loops: Memento detects significant changes in cognitive state from fused EEG, GSR, and PPG signals, selects high-salience frames as in-situ cues, and reports route recall improvements of approximately $K=10$ 9 compared to free recall, together with a $\theta_i\in\{30^\circ,90^\circ,150^\circ,210^\circ,270^\circ,330^\circ\}$ 0 reduction in review time and computational latency $\theta_i\in\{30^\circ,90^\circ,150^\circ,210^\circ,270^\circ,330^\circ\}$ 1 seconds versus $\theta_i\in\{30^\circ,90^\circ,150^\circ,210^\circ,270^\circ,330^\circ\}$ 2 seconds (Ghosh et al., 28 Apr 2025).

6. Evaluation criteria, trade-offs, and open directions

Evaluation protocols vary by domain but consistently test both correctness and efficiency. Embodied-navigation papers report Success Rate, SPL, Progress, PPL, DTG, nDTW, steps to goal, or median path length (Wu et al., 25 Mar 2026, Li et al., 2024, He et al., 24 Jun 2025, Tsybina et al., 16 Apr 2026). Conversational and retrieval systems report accuracy, F1, Recall@5, precision, tool-call accuracy, or LLM-judged performance (Li et al., 27 May 2026, Xu et al., 7 Jul 2026, Dai et al., 14 Apr 2026). Web navigation adds steps per episode, token usage, and runtime (Azam et al., 2 Jun 2025). The repeated presence of path-efficiency and cost metrics indicates that active memory navigation is evaluated not only by whether memory improves end-task success, but by whether it reduces redundant search, unnecessary revisits, or token and compute overhead.

Several trade-offs recur. ReMemNav’s memory-capacity sweep peaks at $\theta_i\in\{30^\circ,90^\circ,150^\circ,210^\circ,270^\circ,330^\circ\}$ 3, with degradation from overly small or overly large buffers (Wu et al., 25 Mar 2026). MemoNav’s forgetting-threshold study shows that easy tasks tolerate higher pruning, whereas harder multi-goal tasks require a moderate threshold (Li et al., 2022). AstraNav-Memory reports that moderate compression rates $\theta_i\in\{30^\circ,90^\circ,150^\circ,210^\circ,270^\circ,330^\circ\}$ 4 maximize the balance between memory efficiency and spatial-semantic fidelity, whereas $\theta_i\in\{30^\circ,90^\circ,150^\circ,210^\circ,270^\circ,330^\circ\}$ 5 causes a sharp success-rate drop (Ren et al., 25 Dec 2025). NaviRAG reports inference latency of about $\theta_i\in\{30^\circ,90^\circ,150^\circ,210^\circ,270^\circ,330^\circ\}$ 6 seconds per query, compared with about $\theta_i\in\{30^\circ,90^\circ,150^\circ,210^\circ,270^\circ,330^\circ\}$ 7 second for vanilla RAG and about $\theta_i\in\{30^\circ,90^\circ,150^\circ,210^\circ,270^\circ,330^\circ\}$ 8 seconds for GraphRAG (Dai et al., 14 Apr 2026). Memento reports that adding GSR and PPG raises computation from $\theta_i\in\{30^\circ,90^\circ,150^\circ,210^\circ,270^\circ,330^\circ\}$ 9 seconds to $0.69/0.56$00 seconds while improving detection accuracy (Ghosh et al., 28 Apr 2025). These results suggest that the central engineering question is selective access at the right granularity, not maximal retention.

A common misconception is that active memory navigation is synonymous with explicit metric mapping. The record does not support that equivalence. SPTM stores no metric information, only connectivity (Savinov et al., 2018). Memory Proxy Maps explicitly remove the need for graphs and odometry (2411.09893). The spiking neuron-astrocyte model biases the agent toward unexplored regions without requiring explicit global statistics and describes its short-term local sensory modulation as a "Topological-Context Memory" (Tsybina et al., 16 Apr 2026). Another misconception is that long context windows alone solve the memory problem. StateLinFormer directly frames fixed context windows as a limitation on persistent memory across extended interactions, while MemCog, NapMem, and NaviRAG all diagnose failures caused by passive invocation, reasoning–retrieval decoupling, structural mismatch, or flat retrieval (Chen et al., 24 Mar 2026, Li et al., 27 May 2026, Xu et al., 7 Jul 2026, Dai et al., 14 Apr 2026).

Open directions are stated explicitly across the literature. MemCog points to adaptive re-structuring of the wiki based on navigation patterns, multi-agent shared memory stores, lighter-weight navigation for cost-sensitive settings, and fully spontaneous recall (Li et al., 27 May 2026). NapMem lists dynamic forgetting/pruning, privacy constraints, multi-user sharing, and scaling studies (Xu et al., 7 Jul 2026). StateLinFormer proposes theoretical analysis of memory-state ergodicity, extensions to other long-sequence domains, and hybrid architectures combining explicit geometric mapping with learned persistent memory (Chen et al., 24 Mar 2026). NaviRAG highlights hybrid horizontal+vertical retrieval, structure-aware organization, and goal-driven memory policies (Dai et al., 14 Apr 2026). Taken together, these directions indicate that active memory navigation is evolving from a navigation-specific technique into a broader design pattern for agents that must decide what to remember, when to retrieve it, how deeply to traverse it, and when to stop.