SimpleMem: Efficient Lifelong Memory System

• SimpleMem is an efficient lifelong memory architecture that minimizes token bloat and redundancy for LLM agents in long-horizon tasks.
• It features a three-stage pipeline—semantic compression, multi-view indexing, and recursive consolidation—to streamline context processing.
• Adaptive, query-aware retrieval dynamically adjusts context assembly, achieving significant token compression and enhanced retrieval precision.

SimpleMem is an efficient lifelong memory architecture for LLM agents operating in complex, long-horizon environments. It is designed to maximize information density and minimize redundant token usage in memory storage and retrieval, directly addressing inefficiencies inherent in both passive long-context extension and costly active iterative reasoning. SimpleMem uses a structured, three-stage pipeline involving semantic lossless compression, multi-view atomic indexing, recursive consolidation into higher-level abstractions, and adaptive, query-aware retrieval, attaining a favorable balance of retrieval precision, compression, and scaling (Liu et al., 5 Jan 2026).

1. Architectural Overview and Motivation

SimpleMem targets the problem that LLM agents incur substantial inefficiency and redundancy when tasked with persistent memory: storing entire interaction histories leads to significant token bloat, while aggressive iterative reasoning to cull irrelevant context increases compute cost. The architecture operates as a loop with three principal stages—semantic structured compression, recursive memory consolidation, and adaptive query-aware retrieval—in which raw multi-turn dialogues are transformed into compact, indexed memory units, abstracted into higher-level representations, and then dynamically retrieved in response to task queries. Each cycle closes with new experience ingestion, completing the loop: ingestion → indexing → consolidation → retrieval → ingestion (Liu et al., 5 Jan 2026).

2. Semantic Structured Compression

The first stage reduces raw interaction transcripts to minimal, self-contained "memory units" via entropy-aware filtering, coreference/temporal normalization, and atomistic segmentation.

Entropy-Aware Filtering

Incoming dialogue is divided into overlapping windows $W_t$ (default: $W=10$ turns, stride=5). For each window, an entropy-based gate quantifies the utility of the content: $$H(W_t) = \alpha\,\frac{|\mathcal E_{\mathrm{new}}|}{|W_t|} + (1-\alpha)\left[1 - \cos(E(W_t), E(H_{\mathrm{prev}}))\right],$$ where $\mathcal E_{\mathrm{new}}$ are new entities, $E(\cdot)$ denotes dense embeddings, and $\alpha$ tunes the balance between entity-level novelty and semantic divergence. Windows with $H(W_t) < \tau_{\mathrm{redundant}}$ (default $\tau=0.35$) are dropped, minimizing the accumulation of redundant or low-salience content.

Memory Unit Segmentation and Normalization

The retained windows are processed by a neural prompt/model function $\mathcal F_\theta$: $$m_k = \mathcal F_\theta(W_t) = \Phi_{\mathrm{time}} \left( \Phi_{\mathrm{coref}}(\Phi_{\mathrm{extract}}(W_t)) \right),$$ where candidate facts are extracted, coreferences resolved, and temporal expressions normalized. The resulting $W=10$0 are context-independent, minimal units annotated with explicit timestamps, entity lists, and optional salience/topic tags.

Multi-View Indexing

Each $W=10$1 is indexed in a tri-view structure comprising:

• Semantic Layer: $W=10$2
• Lexical Layer: $W=10$3
• Symbolic Layer: $W=10$4 (metadata: timestamps, entities)

This enables hybrid retrieval using both fuzzy semantic similarity and exact symbolic or lexical filtering.

3. Recursive Memory Consolidation

To further compress and abstract the active memory store, SimpleMem periodically clusters and merges memory units using semantic and temporal affinity.

Affinity-Based Clustering

Pairwise affinity between units $W=10$5 and $W=10$6 is computed as: $W=10$7 with $W=10$8 (typical 0.7) controlling the weight of semantic vs. temporal proximity, and $W=10$9 (e.g., 0.1) setting the temporal decay rate. Units exceeding the clustering threshold $$H(W_t) = \alpha\,\frac{|\mathcal E_{\mathrm{new}}|}{|W_t|} + (1-\alpha)\left[1 - \cos(E(W_t), E(H_{\mathrm{prev}}))\right],$$0 (default 0.85) form a cluster $$H(W_t) = \alpha\,\frac{|\mathcal E_{\mathrm{new}}|}{|W_t|} + (1-\alpha)\left[1 - \cos(E(W_t), E(H_{\mathrm{prev}}))\right],$$1.

Abstract Representation Synthesis

Each cluster $$H(W_t) = \alpha\,\frac{|\mathcal E_{\mathrm{new}}|}{|W_t|} + (1-\alpha)\left[1 - \cos(E(W_t), E(H_{\mathrm{prev}}))\right],$$2 is synthesized (via $$H(W_t) = \alpha\,\frac{|\mathcal E_{\mathrm{new}}|}{|W_t|} + (1-\alpha)\left[1 - \cos(E(W_t), E(H_{\mathrm{prev}}))\right],$$3) into a single abstract memory $$H(W_t) = \alpha\,\frac{|\mathcal E_{\mathrm{new}}|}{|W_t|} + (1-\alpha)\left[1 - \cos(E(W_t), E(H_{\mathrm{prev}}))\right],$$4: $$H(W_t) = \alpha\,\frac{|\mathcal E_{\mathrm{new}}|}{|W_t|} + (1-\alpha)\left[1 - \cos(E(W_t), E(H_{\mathrm{prev}}))\right],$$5 with original $$H(W_t) = \alpha\,\frac{|\mathcal E_{\mathrm{new}}|}{|W_t|} + (1-\alpha)\left[1 - \cos(E(W_t), E(H_{\mathrm{prev}}))\right],$$6 archived into cold storage. The abstract replaces granular units in the active index, maintaining a compact memory footprint while preserving the ability to recover details as needed. This compression is linear in the number of abstractions and mitigates both redundancy and context inflation.

4. Adaptive Query-Aware Retrieval

At inference time, SimpleMem employs a hybrid, query-sensitive retrieval mechanism to construct the context for LLM prompting.

Hybrid Relevance Scoring

Given query $$H(W_t) = \alpha\,\frac{|\mathcal E_{\mathrm{new}}|}{|W_t|} + (1-\alpha)\left[1 - \cos(E(W_t), E(H_{\mathrm{prev}}))\right],$$7, each candidate $$H(W_t) = \alpha\,\frac{|\mathcal E_{\mathrm{new}}|}{|W_t|} + (1-\alpha)\left[1 - \cos(E(W_t), E(H_{\mathrm{prev}}))\right],$$8 is scored: $$H(W_t) = \alpha\,\frac{|\mathcal E_{\mathrm{new}}|}{|W_t|} + (1-\alpha)\left[1 - \cos(E(W_t), E(H_{\mathrm{prev}}))\right],$$9 where $\mathcal E_{\mathrm{new}}$0 is the query embedding and $\mathcal E_{\mathrm{new}}$1 allows for strict symbolic filters (e.g., timestamps in range).

Query Complexity and Dynamic Retrieval Depth

A small classifier predicts query complexity $\mathcal E_{\mathrm{new}}$2 from query features (length, syntactic complexity, abstraction). The number of candidates retrieved is dynamically adjusted: $\mathcal E_{\mathrm{new}}$3 where $\mathcal E_{\mathrm{new}}$4 (typical values: $\mathcal E_{\mathrm{new}}$5, $\mathcal E_{\mathrm{new}}$6, $\mathcal E_{\mathrm{new}}$7, $\mathcal E_{\mathrm{new}}$8). Simple queries retrieve a minimal number of abstracted units; complex queries expand retrieval to a deeper context window.

Final Context Assembly

The top-$\mathcal E_{\mathrm{new}}$9 ranked units are assembled: $E(\cdot)$0 and prepended to the generation prompt, optimizing context relevance and capacity usage.

5. High-Level Information Flow and System Diagram

The pipeline integrates three operational stages and background consolidation, as summarized in the following block diagram:

$H(W_t) < \tau_{\mathrm{redundant}}$1

6. Key Hyperparameters and Operational Trade-Offs

SimpleMem's effectiveness is governed by several critical hyperparameters:

Stage	Parameter	Typical Value / Function
1	Window size $E(\cdot)$1, stride	$E(\cdot)$2, stride=5
1	Novelty balance $E(\cdot)$3	$E(\cdot)$4
1	Redundancy drop $E(\cdot)$5	$E(\cdot)$6
2	Cluster threshold $E(\cdot)$7	$E(\cdot)$8
2	Semantic-vs-temporal $E(\cdot)$9	$\alpha$0
2	Temporal decay $\alpha$1	$\alpha$2
3	Base retrieval $\alpha$3	$\alpha$4
3	Retrieval bounds $\alpha$5	$\alpha$6
3	Retrieval scale $\alpha$7	$\alpha$8

These parameters instantiate explicit trade-offs: For instance, aggressive filtering (low $\alpha$9) may risk retaining noise, while a high threshold can omit marginally relevant content. Similarly, tuning $H(W_t) < \tau_{\mathrm{redundant}}$0 too low induces over-broad abstraction, whereas excessive strictness fragments memory and loses generalization. Retrieval bounds influence token efficiency versus the risk of omitting critical details (Liu et al., 5 Jan 2026).

7. Performance Characteristics and Applications

Experimental benchmarks demonstrate that SimpleMem consistently achieves superior accuracy, retrieval efficiency, and reduced inference-time token usage relative to baselines, with an average F1 improvement of 26.4% and up to 30-fold inference token compression (Liu et al., 5 Jan 2026). By unifying high-density compression, multi-faceted retrieval, and dynamic context assembly, SimpleMem is suited for LLM agents requiring scalable, lifelong memory with minimal resource overhead, supporting advanced multi-turn reasoning, and complex environment interactions. The codebase is available at https://github.com/aiming-lab/SimpleMem.