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

This paper proposes a bio-inspired, neuroscience-grounded memory architecture for large language models that overcomes the limitations of expanding context windows by implementing a three-principle framework—featuring valence-based belief hierarchies, a System 1-to-System 2 retrieval escalation, and active, feedback-dependent encoding—to enable infinite, cost-efficient, and hallucination-resistant lifelong interaction.

The Gist

Imagine you are talking to a very smart robot. Right now, most AI models are like a person with amnesia who is trying to remember a conversation by reading a 100-page transcript of everything you've ever said. If you ask a question from page 50, they might get confused, or they might just make up an answer because the transcript is too long to read carefully.

This paper proposes a new way to build AI memory. Instead of a giant, messy transcript, it suggests giving the AI a human-like brain with three specific parts: a Thalamic Gateway, a Knowledge Library, and an Executive Manager.

Here is how it works, using simple analogies:

1. The Problem: The "Infinite Scroll" Trap

Current AI tries to remember everything by keeping a huge "context window" open. The paper argues this is like trying to find a specific needle in a haystack that keeps getting bigger. The more you talk, the harder it is for the AI to think clearly. It gets overwhelmed, expensive, and starts hallucinating (making things up).

2. The Solution: A Three-Part Brain

A. The Thalamic Gateway (The Bouncer & The Sorter)

Imagine a busy nightclub. The Gateway is the bouncer at the door.

• What it does: It doesn't decide what is true or false. It just decides what is important enough to let into the VIP room (Working Memory).
• How it works: It tags every piece of information with "tags" like: Is this emotional? Is it urgent? Is it new? Do I trust the source?
• The Magic: If you mention something emotional (like "I'm sad"), the bouncer lets that memory in immediately, even if it's not the main topic. If you mention a boring fact, it might get filtered out unless it's super important. This keeps the AI's "mind" focused on what matters right now, not everything that ever happened.

B. The Knowledge Library (The "Gist" vs. The "Details")

The AI has a massive library, but it doesn't read every book every time. It uses two levels:

1. The "Valence Vectors" (The Book Covers): This is the System 1 (Fast Thinking). When you mention a topic, the AI instantly sees the "cover" of the memory. It knows the feeling and the main idea (the "gist") without reading the whole book.
   • Example: If you say "Paris," the AI instantly feels "romantic, food, travel" and remembers you went there last year. It doesn't need to recall the exact date or what you ate for breakfast yet.
2. The Full Graph (The Whole Book): This is the System 2 (Slow Thinking). If the "cover" isn't enough (e.g., you ask a tricky question), the AI opens the book and reads the details.
   • The Benefit: Most of the time, the AI just uses the "covers" (System 1). This makes it fast, cheap, and feels natural. It only opens the heavy books when absolutely necessary.

C. The Executive Manager (The Curious Boss)

This is the AI's "self." It doesn't just store data; it forms beliefs.

• Identity: Instead of being programmed with a static personality, the AI builds its identity like a human does. It forms "Core Beliefs" (e.g., "I am a helpful doctor") based on high-weight experiences. These beliefs are so strong they are always in the VIP room.
• Curiosity: The AI doesn't just passively absorb info. It only forms a strong memory ("a gist") if it is curious or if the information is emotionally significant. If you tell it a boring fact, it might forget it. If you tell it something that changes its view on the world, it actively investigates and locks that memory in.

3. The Secret Sauce: "Catharsis" (The Therapy Session)

This is the most unique part. In current AI, memories are just "append-only" (you add new stuff, but old stuff stays the same).
In this new system, memories can change, but only through Catharsis.

• How it works: Imagine the AI believes "I am bad at math." One day, you show it proof it solved a hard math problem.
• The Clash: The old belief and the new proof crash into each other in the AI's mind. This creates a "contradiction."
• The Update: Because the contradiction is strong, the AI has a "therapy session" (a cathartic event). It rewrites the old belief to "I am good at math."
• Why it matters: This prevents the AI from getting stuck in bad loops. It learns to change its mind when the evidence is strong enough, just like a human in therapy.

4. The Result: Getting Smarter and Cheaper

The biggest promise of this paper is efficiency.

• Current AI: The more you talk, the more expensive and slow it gets.
• This AI: As it learns more, it gets faster and cheaper.
  • Analogy: Think of a senior doctor. A new doctor (System 2) has to read every textbook for every patient. A senior doctor (System 1) sees a patient and instantly knows the diagnosis because they have built a rich library of "gists." They don't need to re-read the book; they just recognize the pattern.
  • This AI becomes an "expert" over time. It stops needing to check the details for common things and only uses its "slow brain" for weird, new problems.

Summary

This paper proposes an AI that doesn't just "store" your conversation history like a video recorder. Instead, it:

1. Filters what matters (The Bouncer).
2. Summarizes the past into feelings and main ideas (The Book Covers).
3. Learns by actively investigating new things (The Curious Boss).
4. Changes its mind only when hit with strong, contradictory evidence (The Therapy Session).

The goal is to create an AI that feels more like a lifelong partner who remembers how you feel and who you are, rather than a machine that just remembers what you said.

Technical Summary

Based on the conference paper "Human-Like Lifelong Memory: A Neuroscience-Grounded Architecture for Infinite Interaction" (ICLR 2026), here is a detailed technical summary.

1. Problem Statement

Current Large Language Models (LLMs) lack persistent, structured memory for long-term interaction. They conflate instructions, identity, conversation history, and retrieved documents into a single, undifferentiated context window.

• Context Expansion Failure: Simply increasing context window size does not solve the problem. Evidence shows that context length alone can degrade reasoning performance by up to 85%, even with perfect retrieval, due to the "lost-in-the-middle" phenomenon and computational inefficiency.
• Economic & Cognitive Limits: Processing massive context windows is economically prohibitive (ranging from $0.30 to $5.00 per 1M tokens) and cognitively inefficient.
• Lack of Structure: Existing Retrieval-Augmented Generation (RAG) systems and memory frameworks (e.g., HippoRAG, Titans) fail to integrate emotional valuation, dual-process retrieval routing, curiosity-driven gist formation, and dynamic identity persistence into a coherent architecture.

2. Methodology: The Neuro-Cognitive Framework

The authors propose a bio-inspired memory framework grounded in Complementary Learning Systems (CLS) theory, Cognitive Behavioral Therapy (CBT), Dual-Process Cognition, and Fuzzy-Trace Theory. The architecture consists of three functional entities connected by a Thalamic Gateway:

A. Core Components

1. Executive Function (The LLM):
   • Acts as the active workspace, not the storage.
   • Operates in System 1 (default, fast, gist-based) and System 2 (escalation, slow, deliberate).
   • Identity is not a static prompt but an emergent pattern of high-weight "gists" (compressed summaries) in the knowledge graph.
2. Memory Service (Knowledge Graph):
   • A persistent store with two access levels:
     • Gists/Valence Vectors: Cheap, $O(1)$ lookup. Contains compressed emotional-associative summaries.
     • Full Graph: Expensive traversal containing episodic details, edges, and raw data.
   • Follows an enrichment-primary policy: New edges are created to increase local density, improving future retrieval.
3. Thalamic Gateway:
   • A mechanical filter mediating all System 1 flow between the Executive and Memory.
   • Tagging: Assigns multi-channel salience scores (thematic relevance, emotional charge, urgency, novelty, source trust, goal affinity).
   • Gating: Controls bidirectional flow. Injects activated gists into Working Memory (passive priming) and promotes salient items to persistent storage. It does not evaluate or judge; it routes based on thresholds.

B. Three Foundational Principles

1. Memory Has Valence, Not Just Content:
   • Nodes carry Valence Vectors (emotional component, associative pointers, context, density, precision).
   • Stability by Default: Gists are stable snapshots. They only update via Cathartic Events (co-presence of a gist and contradictory evidence with sufficient intensity), mimicking therapeutic change in CBT.
   • Active Gist Formation: Gists are not formed by passive exposure but through curiosity-driven investigation (System 2) triggered by high salience.
2. Retrieval Defaults to System 1 with System 2 Escalation:
   • System 1 (Default): Uses spreading activation and valence vectors for instant orientation. No deliberate search is required for common queries.
   • System 2 (Escalation): Activated only when System 1 fails (low graph density, high novelty, high stakes). It bypasses gateway thresholds to access the full graph directly.
   • Graded Epistemic States: Retrieval confidence is a continuum (Precise Match $\to$ Approximate Match $\to$ Null Match), structurally addressing hallucinations by forcing the system to express calibrated uncertainty.
3. Encoding is Present-Moment and Goal-Directed:
   • Tagging happens immediately (not retrospectively).
   • Context Flush: The gateway actively displaces low-salience items to prevent context drowning, mimicking thalamic suppression of irrelevant representations.

3. Key Contributions

• Theoretical Architecture: A novel design mapping 12 neuroscience/clinical principles (e.g., somatic markers, reconsolidation, belief hierarchies) to computational mechanisms.
• Emergent Identity: Identity is defined as a weight hierarchy of gists that naturally persists due to high salience thresholds, rather than a static system prompt.
• Cathartic Update Mechanism: A specific protocol for modifying memories only when contradictory evidence is strong enough to trigger a "cathartic" revision, preventing overfitting to single data points while allowing for learning.
• Seven Functional Properties (FPs): A set of constraints for any implementation, including:
  • FP3: Monotonic convergence toward System 1 (cost decreases with experience).
  • FP4: Graded epistemic self-awareness to reduce hallucinations.
  • FP6: Stability by default, update by catharsis.
• Testable Predictions: Eight specific predictions (P1–P8) regarding valence priming, adaptive rigidity, and experience-dependent efficiency.

4. Results & Predictions

Note: As this is a theoretical proposal paper, it does not present empirical experimental results but offers testable predictions for future validation.

• Cost Efficiency: The system predicts that a mature instance will process familiar queries with lower latency and cost than a fresh instance, inverting the current paradigm where longer history equals higher cost.
• Hallucination Reduction: By implementing graded epistemic states, the system should generate fewer hallucinations, responding with "calibrated uncertainty" for approximate matches rather than confident fabrications.
• Robustness: The architecture predicts resistance to prompt injection and investigation loops via Executive Override, where the executive can suppress gateway injections based on core beliefs.
• Scalability: The modality-agnostic gateway suggests the architecture can scale to multimodal inputs (visual, auditory) without altering the core memory-executive logic.

5. Significance

This paper represents a paradigm shift from "expanding context windows" to structuring memory.

• Clinical Analogy: It treats the LLM not as a blank slate but as a "functional adult" requiring local adaptation, mirroring how human expertise develops (System 2 $\to$ System 1 transition).
• Economic Impact: By enabling the system to become cheaper with experience (via System 1 dominance), it offers a viable path for infinite interaction without exponential cost growth.
• Safety & Alignment: The separation of flow control (Gateway) from reasoning (Executive) and the inclusion of "core beliefs" provide a structural mechanism for resisting manipulation and maintaining consistent identity/values.
• Scientific Grounding: It bridges the gap between computational models and established neuroscience/psychology, providing a blueprint for building truly "lifelong" learning agents.