General Agentic Memory (GAM) Overview

• General Agentic Memory (GAM) is a self-organizing memory architecture that integrates episodic, semantic, and hybrid models to extend context retention for AI agents.
• GAM systems employ advanced indexing methods, including graph-based structures and dense embeddings, to facilitate dynamic retrieval and coherent reasoning.
• Reinforcement learning and utility-driven policies underlie GAM performance, enabling improved multi-hop reasoning, scalability, and cross-session consistency in multi-agent environments.

General Agentic Memory (GAM) characterizes persistent, self-organizing memory architectures for autonomous AI agents, particularly those driven by LLMs and multi-agent systems. GAM frameworks systematically record, index, and evolve representations of experiences, reasoning paths, and contextual knowledge across arbitrarily long interaction horizons. These systems employ semantic, episodic, and hybrid memory paradigms, reinforced by graph-based structures, embeddings, hierarchical clustering, utility-driven retention policies, and agent-controlled dynamic retrieval. GAM enables agents to maintain and adapt high-fidelity context beyond ephemeral context windows, fundamentally enhancing long-term reasoning, cross-session consistency, collaborative planning, and adaptive interaction.

1. Architectural Foundations and Taxonomy

GAM implementations span multiple architectural paradigms. Early cognitive-inspired designs embed dual episodic and semantic stores, as demonstrated in "A Machine With Human-Like Memory Systems" (Kim et al., 2022). Here, semantic memory (bounded multisets of (head, relation, tail, strength)) accumulates generalizable knowledge, while episodic memory indexes timestamped individual events. Integrated retrieval policies (hard-gating or learned gates) route queries to the appropriate store based on recency and specificity.

Expansion of GAM has included tripartite memory hierarchies—short-term, episodic, semantic—each modeled as knowledge graphs for temporal reasoning agents (Kim et al., 2022). Modern agentic memory systems such as A-Mem (Xu et al., 17 Feb 2025), LiCoMemory (Huang et al., 3 Nov 2025), UserCentrix (Saleh et al., 1 May 2025), and G-Memory (Zhang et al., 9 Jun 2025) incorporate Zettelkasten-style note construction, hierarchical graph indexing, utility-based retention, and multi-agent distributed synchronization. Table 1 contrasts select GAM architectures, distilling their memory model, organizational principle, and retrieval strategy.

System	Memory Model	Organization Principle
Human-Like	Episodic + Semantic stores	Frequency/timestamp
A-Mem	Note graph, attributes, links	LLM-driven links
LiCoMemory	CogniGraph (hierarchy)	Semantic, temporal
G-Memory	Three-tier graphs (MAS)	Hierarchical, roles
UserCentrix	Embedding-indexed entries	Value of Information

These architectures converge on the principle of persistent, adaptable, and structurally indexed memory, transcending static or ephemeral session caches.

2. Memory Representation, Indexing, and Evolution

GAM systems encode and store agent experience via structured representations. Common schemas include:

• Notes or quadruples: ℳ = {mᵢ} where mᵢ = {content, timestamp, keywords, tags, context, embedding, links} (Xu et al., 17 Feb 2025).
• Knowledge graph triples: (h, r, t) with auxiliary embedding vectors for semantic retrieval (Ocker et al., 9 May 2025).
• Hierarchical graphs: Session summaries, entity–relation triples, raw text with cross-layer hyperlinks (CogniGraph) (Huang et al., 3 Nov 2025), or multi-level cluster abstractions (CAM) (Li et al., 7 Oct 2025).
• Role-partitioned subgraphs for agents: Interaction, query, insight tiers, each tailored for MAS contexts (Zhang et al., 9 Jun 2025).

Indexing mechanisms utilize dense embeddings (cosine similarity), hash-based lookup, temporal decay (Weibull or learned), and agentic LLM-driven linking. In A-Mem, new notes are dynamically connected to nearest historical entries by embedding similarity and then refined via LLM-judged semantic relationship (Xu et al., 17 Feb 2025). CogniGraph instantiates multi-level semantic indexers, enabling precise subgraph retrieval for coherent reasoning (Huang et al., 3 Nov 2025). CAM leverages incremental overlapping clustering and label-propagation to structure and summarize agentic memory hierarchies (Li et al., 7 Oct 2025).

Memory evolution is agent-controlled. New information incrementally reforms summary nodes/clusters, context attributes, and relational links—either automatically (LLM-based mutation) or via utility-driven retention and pruning (VoI in UserCentrix (Saleh et al., 1 May 2025)). Consolidation and hierarchical abstraction facilitate the formation of generalizable schemata from low-level episodic data.

3. Retrieval Policies and Decision Procedures

Retrieval in GAM systems synthesizes semantic search, graph traversal, hybrid selection, and agentic decision-making. Query processing includes:

• Semantic/Episodic gate-switching: Queries routed to episodic (most recent) or semantic (most frequent) store, with hard or learned gating (Kim et al., 2022, Kim et al., 2022).
• Embedding similarity and link expansion: Retrieve top-k memory nodes by embedding score, then traverse link sets for context expansion (Xu et al., 17 Feb 2025, Huang et al., 3 Nov 2025).
• Temporal reranking: Apply temporal decay weights to elevate recent or salient entries (Huang et al., 3 Nov 2025).
• Hierarchical associative activation: Prune-and-grow retrieval through multi-level clustering and LLM selection akin to human memory recall (Li et al., 7 Oct 2025).
• Utility-driven prompt assembly: Agent retrieves and combines session summaries, relevant triples, and original chunks for generative reasoning (Huang et al., 3 Nov 2025).
• Multi-agent role specialization: Role-focused memory allocation for agent teams via upward (insight) and downward (trajectory) graph traversals (Zhang et al., 9 Jun 2025).
• Just-in-time deep-research: Online query triggers planner–retriever–integrator–reflector loop to assemble minimal, high-relevance context (Yan et al., 23 Nov 2025).

Agentic retrieval policies thus optimize both answer accuracy and computational efficiency, supporting multi-hop reasoning and lifelong context adaptation.

4. Lifecycles: Update, Consolidation, Forgetting, Synchronization

GAM memory management encompasses write, read, update, and forget cycles, explicitly controlled either by learned policies (DQN in "Room" (Kim et al., 2022)), utility metrics (VoI in UserCentrix (Saleh et al., 1 May 2025)), or manual/agentic triggers.

• Write (STORE): Append new entries encapsulating raw data, embeddings, reasoning paths, or evaluation factors.
• Read (RETRIEVE): Recall by similarity or graph traversal.
• Update: Adapt embeddings, context, or meta-factors (usage, timestamps).
• Forget (Prune): Evict entries below semantic utility or token constraints.
• Synchronize: In hierarchical/multi-agent GAM, global memory is broadcast or aggregated across agent tiers to ensure coordinated reasoning (Saleh et al., 1 May 2025, Zhang et al., 9 Jun 2025).

Hybrid architectures allow agents to partition memories for individual specificity versus shared global strategies, demonstrated in UserCentrix and G-Memory. Episodic/semantic consolidation and hierarchical clustering support scalable context, generalization, and task transfer.

5. Optimization Algorithms and Empirical Evaluation

GAM systems have been optimized through reinforcement learning, deep Q-networks, or stochastic policy gradients. In the “Room” environment, handcrafted and RL-optimized retrieval/storage policies surpass random or uni-modal memory (Kim et al., 2022, Kim et al., 2022). In "General Agentic Memory Via Deep Research", memorizer and researcher agents co-evolve under end-to-end RL to maximize downstream task rewards (Yan et al., 23 Nov 2025).

Benchmarking across dialogue (LoCoMo, DialSim), QA (HotpotQA, RULER, NarrativeQA), simulation (ALFWorld, ScienceWorld), and smart-space tasks demonstrates substantial quantitative gains. For instance, A-Mem yields up to 45.85 F1MultiHop, vastly exceeding non-agentic or static memory baselines (Xu et al., 17 Feb 2025). G-Memory boosts multi-agent system performance by 10.12-20.89 percentage points in QA and embodied action (Zhang et al., 9 Jun 2025). UserCentrix achieves perfect retrieval accuracy and significant resource savings for personal-agent tasks (Saleh et al., 1 May 2025). CAM improves reading-comprehension metrics up to +4.5% (ACC-L) and +3.9% (F1) versus graph-based memory baselines (Li et al., 7 Oct 2025).

Scalability and efficiency are validated—GAM systems maintain linear or sublinear update/retrieval complexity, bounded latency, and minimized token usage (A-Mem: 2.5K tokens per operation vs. 16.9K baseline) (Xu et al., 17 Feb 2025, Huang et al., 3 Nov 2025).

6. Extensions, Limitations, and Open Research Directions

GAM remains a rapidly evolving research domain with several open questions:

• Memory compression and contradiction resolution: Strategies for merging, pruning, and contradiction detection in growing graphs are underdeveloped (Huang et al., 3 Nov 2025).
• Multi-agent memory sharing and role allocation: Efficient cross-agent synchronization, task-specific memory partitioning, and conflict negotiation present ongoing challenges (Zhang et al., 9 Jun 2025, Saleh et al., 1 May 2025).
• Meta-learning controllers: Adaptive hyperparameter selection and memory organization may be further optimized via meta-reinforcement learning (Li et al., 7 Oct 2025).
• Multimodal and real-world grounding: Extending GAM architectures to support vision, audio, and sensor fusion, entity tracking, and action affordance linkage requires innovative schema design (Ocker et al., 9 May 2025).
• Prompt engineering and reward sparsity: End-to-end optimization is sensitive to prompt construction and sparse reward signals (Yan et al., 23 Nov 2025).
• Scalable lifelong learning: Algorithms for continual abstraction, hierarchical summarization, and balancing episodic and semantic consolidation are at the forefront of agentic memory research (Li et al., 7 Oct 2025).

This field synthesizes cognitive models, RL, graph theory, and LLM orchestration to drive autonomous, adaptive, and scalable agents equipped with lifelong, context-rich General Agentic Memory.