Publication details
About this article
- Author
- Prof Amy Shi-Nash and Dr Jizhizi Li
- Publication date
- 29 May 2026
- Classification
- NxtLab research article
Research overview
Overview
We design memory-augmented architectural layers for personalised and long-lived human-centric AI systems.
The objective is to develop AI systems that convert user input into personalized knowledge representations that are portable across computational environments, persistent over time, and privately maintained at individual level, while becoming progressively more data-efficient—requiring fewer generic prompts and less new information over time as structured understanding accumulates.
Memory architecture
Working context
The live conversation: high-precision, session-level detail used to reason in the moment.
Session summary
A structured compression of the interaction, created so it can be recalled later.
Consolidation
Evidence is aggregated, conflicts detected and confidence adjusted before anything becomes durable.
Persistent knowledge
Preferences, facts and stable relationships held as updateable, queryable long-term memory.
Retrieval
Relevant memory is selected and projected back into the working context for the next interaction.
Retrieval returns selected memory to working context, so the next interaction starts with what is already known rather than from zero.
Graph memory
Semantic structure: entities, attributes, relations, provenance and evidence. Used for structured reasoning and stable long-horizon consistency.
Episodic memory
Context-rich traces held for similarity matching, fuzzy recall and nuanced retrieval. Complements the graph rather than replacing it.
Consolidation links the two: episodic evidence can become semantic knowledge, while reconsolidation updates that knowledge when memory is reactivated.
Read the diagram as text
Working context → session summary → consolidation → persistent knowledge → retrieval, with retrieval returning to working context. Persistent knowledge is represented twice: as graph memory for structured entities, relations and provenance, and as episodic memory for context-rich traces and similarity recall. Consolidation moves supported evidence from episodic traces into semantic memory; reconsolidation revises both when new evidence arrives.
1. Temporal Memory Architecture
Memory is organised across time: what stays in the working context, what is summarised after each interactive session, and what becomes durable knowledge.
This is conceptually aligned with hippocampal–cortical consolidation — fast encoding followed by gradual stabilisation.
1.1 Short-term — Working Context
Session-level information used directly during interaction for high-fidelity reasoning.
- Context window grounding
- High precision, non-persistent
- Optimised for real-time continuity
1.2 Medium-term — Session Summary
Structured compression of each interactive session for scalable recall across time.
- Compression for long-horizon scale
- Stable cross-session recall
- Bridge into consolidation
1.3 Long-term — Persistent Knowledge
Durable memory of preferences, facts, and stable relationships that can evolve over time.
- Queryable and updateable
- Supports long-lived personalisation
- Designed for continuous evolution
2. Structural Memory Representation
We treat long-term memory as an evolving representation, not a static log.
Our system combines semantic structure (graph) with episodic traces (vector retrieval), supported by conflict-aware updates.
2.1 Semantic — Graph-based Memory
A knowledge graph of entities, attributes, relations, provenance, and evidence.
- Structured reasoning and querying
- Provenance and evidence links
- Stable long-horizon consistency
2.2 Episodic — Vector-based Memory
Context-rich traces for fuzzy recall, similarity matching, and nuanced retrieval.
- High recall for soft memories
- Similarity-based retrieval
- Complements the knowledge graph
2.3 Safety and Stability — Privacy, Conflict, Confidence, Versioning
Updates are cautious: conflicts trigger confidence changes and versioned states when needed.
- Conflict detection across sessions
- Confidence tracking and recency
- Version control for memory evolution
3. System-level Memory Capabilities
Beyond storage, we study what memory enables at a system level: portability, separation by context, and cross-modal consolidation.
This is where our unified multi-modal direction sits, including collaboration with partners.
3.1 Portability — Portable Memory
Exploring representations and protocols that allow memory to move across products, business contexts, and heterogeneous agent ecosystems.
- Interoperable memory formats
- Policy-aware transfer
- Ownership and governance by design
- Extensible to third-party agents and skill systems
3.2 Separation — Splittable Memory
Context-aware activation that separates work, personal, and other domains without leakage.
- Selective memory projection
- Role/persona separation
- Reduced privacy bleed
3.3 Multimodal — Cross-modal Consolidation
Unifying signals across text and other modalities into shared memory objects with evidence and confidence.
- Unified memory objects across modalities
- Cross-modal reinforcement for robustness
- Conflict-aware multi-modal updating
Outcome
Temporal architecture, structural representation, and system-level capabilities together form a persistent and evolving memory substrate.
Over time, this supports consistent long-term identity grounded in consolidated and personalised knowledge rather than transient interaction history.
Research collaboration
Interested in collaboration?
We work with research partners, enterprise teams, and product builders on:
- Memory evaluation
- Consolidation mechanisms
- Unified multi-modal memory systems
Contact us to explore research collaborations.
Discuss a collaboration- memory
- memory-augmented AI
- AI architectures
- research
- temporal memory
- structural memory representation
- system-level memory
- multimodal