Research
How should an AI remember, and what should it forget?
Pre-doctoral research on episodic and semantic memory in AI systems, grounded in cognitive neuroscience and Complementary Learning Systems theory, and tested in a running city of AI agents.
The question
How can AI systems build and keep memory that supports a persistent identity, continuous learning and genuine episodic recall?
The field is moving fast. Letta (formerly MemGPT) introduced tiered memory with core, recall and archival stores, open-source agents like Moltbot carry memory across sessions, and the December 2025 survey "Memory in the Age of AI Agents" became Hugging Face's top daily paper.
Most of that work is semantic memory and document retrieval. True episodic memory, recalling a specific experience grounded in time and place, remembering not just what happened but what it was like, is still open. My work asks what cognitive neuroscience can contribute to closing that gap.
The testbed
Pantheon: memory tested against behaviour
The AI agents that run my portfolio live in a small city when their work is done. Their memory is the experiment: give them a memory, let them live, and measure what forms.
Four kinds of memory, kept apart
Facts a resident started with, things that happened to it, things it heard, and its own current interpretation. A rumour never becomes a fact, and an opinion can change.
Strength, not a cap
Every memory has a strength. Use strengthens it, time fades it, and facts never fade. There is no fixed number of memories a resident may keep.
Rest is consolidation
Rest replays recent experience. What is reinforced stays in working memory; what falls below a threshold moves to an archive.
Forgotten is not gone
A cue that resembles an archived memory can bring it back, so forgetting is a change in reach rather than a deletion.
The instrument is replay. Every decision is logged with the memory the resident had in hand, so a day can be run again with a different memory.
How Pantheon worksThe current landscape
What exists already, and where the open problems are.
What exists now
- Virtual context management (MemGPT/Letta) with hierarchical memory tiers
- Persistent agent memory for cross-session continuity (Moltbot, Letta)
- RAG systems for retrieval-augmented generation from external knowledge
- Self-directed memory editing where LLMs manage their own memory through tool use
- MemRL framework (January 2026) for self-evolving agents via RL on episodic memory
Open questions
- True episodic recall: remembering experiences, not just retrieving documents
- Catastrophic forgetting: in continuous learning without replay
- Memory consolidation: integrating new experience without destabilizing what is known
- Temporal encoding: representing when something happened, not just what
- Identity persistence: keeping a coherent self across experience and time
Neuroscience foundations
The brain solves the continuous learning problem AI still struggles with. Complementary Learning Systems theory (McClelland, McNaughton and O'Reilly) describes two systems working together.
The hippocampus learns fast: sparse, pattern-separated records of specific experiences. The neocortex learns slowly, extracting regularities into semantic knowledge. The hippocampus acts as an index that reactivates distributed traces, and through consolidation those memories come to be held by the neocortex.
Sleep replay is central. Research from UC San Diego and UC Irvine (2025) showed slow-wave sleep interleaving replay of familiar and new traces within single slow waves, so new memories integrate without catastrophic interference. That interleaving principle carries straight over to AI architectures.
Episodic memory
Specific experiences grounded in time and place: not just that something happened, but the experience of it.
For AI: remembering particular interactions and how an understanding formed.
Semantic memory
General knowledge separated from when it was learned. You know Paris is in France without remembering learning it.
For AI: LLMs are strong here. The gap is tracing how an understanding developed.
Procedural memory
Skill built through practice: riding a bike, playing an instrument, a familiar route.
For AI: held in model weights, and hard to update without retraining.
Bio-inspired approaches
- Hippocampal-Augmented Memory Integration (HAMI)
2025. Symbolic indexing, hierarchical memory refinement and structured episodic retrieval, modelled on how the hippocampus operates.
- Sleep Replay Consolidation (SRC)
A sleep-like phase in which a network replays and consolidates memories, allowing continual learning without catastrophic forgetting. Spiking networks are especially promising here.
- Nested Learning
Google Research, NeurIPS 2025. Models as nested learning systems whose parts update at different rates, a "continuum memory system" analogous to hippocampal and neocortical interaction.
Research questions
CLS-inspired architectures
What is the computational analogue of hippocampal and neocortical interaction, and how does it shape a practical architecture?
Replay and consolidation
Can replay during rest allow continuous learning without catastrophic forgetting? What should be replayed, and how often?
Tested in PantheonTemporal encoding
How can a system represent when an experience happened, and what role does that play in recall?
Identity persistence
Which mechanisms keep a self coherent over time, and how does accumulated experience shape it?
Tested in PantheonBeyond RAG
What separates remembering from searching? How can retrieval support recall of experiences rather than document lookup?
Tested in PantheonAuthenticity verification
As synthesis improves, which markers of identity can still show where content came from?
The products ask the same questions
Building real systems shows constraints theory alone does not.
Ajax Studio
Creative persistence: characters and voices that keep their identity from scene to scene and project to project.
VoiceGuard AI
Identity from the security side: what verifies a speaker once any voice can be synthesized?
CacheX
Grounding: an AI that has to keep what it has seen anchored to a market that keeps moving.
Latest insights
All insights- 2026-10-12Memory Consolidation in AI: Learning Without Losing the DetailStoring every episode preserves raw context but leaves the semantic model static, while aggressive summarization updates the model at the cost of losing detailed episodic traces. Complementary Learning Systems theory offers a framework for interleaved replay that could balance these extremes, but practical replay mechanisms introduce bias and scalability challenges.
- 2026-10-08Exploring Episodic Memory in MemCo: Local vs Global Memory Trade-offsMemCo introduces a two‑tier memory system that separates environment‑specific local traces from a global repository of transferable workflows. The paper claims this design improves task success and cuts redundant exploration, but leaves open how episodic recall and consolidation are actually realized.
- 2026-10-05Why Vector Magnitude Matters for Episodic Memory in AI SystemsI explore how treating vectors merely as semantic embeddings overlooks their magnitude and direction, which are crucial for modeling hippocampal‑like episodic recall. Preserving these geometric properties lets AI memory‑augmented agents capture memory strength, context, and pattern‑completion dynamics.
Interested in collaboration?
I welcome conversations with researchers, institutions and organizations working on AI memory, Complementary Learning Systems or related areas.