AuthorsDiego C. Lerma-Torres
TL;DR
Human-Like Lifelong Memory uses a thalamic gateway plus valence-vector knowledge graph to turn long interaction histories into cheaper, System 1–style processing over time.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Du et al. (2025) show that context length alone degrades performance by up to 85%, even when retrieval is perfect.
LLMs that stuff everything into one context window lose structure, causing brittle long-term interaction and expensive, unreliable reasoning over long histories.
HOW IT WORKS
Human-Like Lifelong Memory centers on Executive Function and Working Memory, a Memory Service Knowledge Graph, and a Thalamic Gateway that tags, gates, and routes information.
You can think of Human-Like Lifelong Memory like a brain-inspired split between RAM and disk, with a hippocampal-style graph and a thalamus-like router in front.
This architecture lets Human-Like Lifelong Memory use valence vectors and dual System 1/System 2 routing to gain identity, graded confidence, and cheaper expertise that a plain context window cannot achieve.
DIAGRAM
This diagram shows how Human-Like Lifelong Memory routes a single query through System 1 automatic retrieval, then escalates to System 2 deliberate search when needed.
DIAGRAM
This diagram shows how Human-Like Lifelong Memory’s seven functional properties organize around working memory, the gateway, and the knowledge graph.
PROCESS
Executive Function and Working Memory
Human-Like Lifelong Memory uses Executive Function and Working Memory as a capacity limited workspace where high weight gists persist as emergent identity and current context.
Thalamic Gateway Tagging and Gating
The Thalamic Gateway tags each segment with multi channel salience and decides which items to promote to the Memory Service Knowledge Graph or inject back as priming.
System 1 and System 2 Routing
Human-Like Lifelong Memory defaults to System 1 automatic retrieval from gists and escalates to System 2 deliberate search over the full graph when novelty, stakes, or sparsity demand it.
Cathartic Update and Active Gist Formation
Through curiosity driven investigation, Human-Like Lifelong Memory forms new gists and uses cathartic update to modify existing gists when strong contradictions and precision thresholds are met.
KEY CONTRIBUTIONS
Neuroscience grounded dual store architecture
Human-Like Lifelong Memory formalizes a dual store system with Executive Function and Working Memory plus a Memory Service Knowledge Graph, mirroring complementary learning systems.
Valence vectors and belief hierarchy
Human-Like Lifelong Memory attaches valence vectors with precision scalars to gists and maps Beck’s CBT belief hierarchy onto an emergent weight distribution in the knowledge graph.
Seven functional properties and predictions
Human-Like Lifelong Memory specifies seven functional properties, including graded epistemic self awareness and monotonic convergence toward System 1, plus eight concrete experimental predictions.
RESULTS
Context length degradation
85% performance drop
Du et al. (2025) vs short context
Context window pricing
$0.30 to $5.00 per 1M tokens
Gemini 2.5 Flash vs Claude Opus 4.6
Salience channels
6 tagging dimensions
thematic, emotional, urgency, novelty, trust, goal
Functional properties
7 required properties
architecture constraints for implementations
Human-Like Lifelong Memory is motivated by Du et al. (2025), where long contexts cause up to 85% performance degradation even with perfect retrieval. The framework argues that explicit lifelong memory with valence vectors and gating is needed to avoid paying $0.30–$5.00 per million tokens while still maintaining reasoning quality.
BENCHMARK
Human-Like Lifelong Memory is motivated by Du et al. (2025), where long contexts cause up to 85% performance degradation even with perfect retrieval. The framework argues that explicit lifelong memory with valence vectors and gating is needed to avoid paying $0.30–$5.00 per million tokens while still maintaining reasoning quality.
BENCHMARK
Dollar cost per 1M input tokens for different LLMs, motivating Human-Like Lifelong Memory’s emphasis on cheaper System 1 processing.
KEY INSIGHT
Du et al. (2025) show that simply increasing context length can degrade reasoning performance by up to 85%, even with perfect retrieval.
This breaks the common assumption that more context is always better and reveals that undifferentiated long windows can actively harm reasoning instead of helping.
WHY IT MATTERS
Human-Like Lifelong Memory unlocks agents that develop stable identities, graded epistemic confidence, and cheaper expert like System 1 behavior over long lived interactions.
Builders can now design memory systems around valence vectors, belief hierarchies, and thalamic style gateways instead of just scaling context windows and naive retrieval.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Yi Yu, Liuyi Yao et al.
arXiv 2026 · 2026
Agentic Memory (AgeMem) exposes memory management tools, a three-stage progressive RL strategy, and step-wise GRPO directly inside the agent policy to jointly control long-term and short-term memory. On Qwen3-4B-Instruct, AgeMem attains 54.31% average performance across ALFWorld, SciWorld, PDDL, BabyAI, and HotpotQA, exceeding the best baseline A-Mem at 45.74%.
Can Lv, Heng Chang et al.
· 2026
All-Mem organizes long-term agent memory through Online/Offline Decoupling, Agentic Topology Consolidation, and Topology-Aware Retrieval over a topology-structured memory bank. On LoCoMo, All-Mem reaches 54.63 4o-J versus Mem0’s 48.91, and on LongMemEval-S All-Mem reaches 60.20 4o-J versus Mem0’s 55.80.
Answers use this explainer on Memory Papers.
Checking…