Human-Like Lifelong Memory: A Neuroscience-Grounded Architecture for Infinite Interaction

arXiv cs.AI / 4/1/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper argues that simply increasing LLM context windows cannot provide reliable long-term, context-sensitive memory, because context length can substantially degrade reasoning even with perfect retrieval.
It proposes a neuroscience- and cognition-grounded memory architecture for “infinite interaction,” using precomputed emotional-associative “valence vectors” organized in an emergent belief hierarchy.
The framework specifies retrieval behavior as defaulting to fast, automatic System 1-style activation (with System 2-style deliberate retrieval only when necessary) and introduces graded epistemic states to structurally mitigate hallucinations.
It describes active, feedback-dependent encoding via a “thalamic gateway” that routes information between memory stores and an executive process that forms gists through curiosity-driven investigation.
Seven functional properties are outlined as implementation requirements, with the intended outcome that interaction becomes cheaper over time as the system converges toward expertise-like processing.

Abstract

Large language models lack persistent, structured memory for long-term interaction and context-sensitive retrieval. Expanding context windows does not solve this: recent evidence shows that context length alone degrades reasoning by up to 85% - even with perfect retrieval. We propose a bio-inspired memory framework grounded in complementary learning systems theory, cognitive behavioral therapy's belief hierarchy, dual-process cognition, and fuzzy-trace theory, organized around three principles: (1) Memory has valence, not just content - pre-computed emotional-associative summaries (valence vectors) organized in an emergent belief hierarchy inspired by Beck's cognitive model enable instant orientation before deliberation; (2) Retrieval defaults to System 1 with System 2 escalation - automatic spreading activation and passive priming as default, with deliberate retrieval only when needed, and graded epistemic states that address hallucination structurally; and (3) Encoding is active, present, and feedback-dependent - a thalamic gateway tags and routes information between stores, while the executive forms gists through curiosity-driven investigation, not passive exposure. Seven functional properties specify what any implementation must satisfy. Over time, the system converges toward System 1 processing - the computational analog of clinical expertise - producing interactions that become cheaper, not more expensive, with experience.