Total Recall
AI for Lifelong Memory
This workshop focuses on lifelong, evolving memory at the scale of terabytes to petabytes of artifacts per coherent context, corresponding to decades of human digital activity.
Registration opens soon
This workshop focuses on lifelong, evolving memory at the scale of terabytes to petabytes of artifacts per coherent context, corresponding to decades of human digital activity.
Registration opens soon This workshop focuses on lifelong, evolving memory at the scale of terabytes to petabytes of artifacts per coherent context, corresponding to decades of human digital activity. How to represent evolving memory at the human lifelong scale is a vital gap precluding current AI from achieving key aspects of human-level intelligence. AI that can handle a lifetime of memories also promises to alleviate devastating conditions of memory loss and even augment human memory.
Memory at the human lifelong scale extends well beyond today's chatbot conversations, agent traces, and agent memory benchmarks, and poses particular challenges: multiscale consolidation of experience, preserving details through repeated consolidations, retaining rarely used evidence, distinguishing similar events, handling evolving facts and preferences, and cost-efficient encoding, updates, and recall. AI that augments human memory should improve assisted recollection, reduce effort in organizing documents and files, and surface forgotten memories that inform present decisions.
By convening experts in agentic memory, context management, long-context models, continual and lifelong learning, personalized AI, human–computer interaction, and the cognitive neuroscience of memory, we aim to shape a coherent framework and design adequate benchmarks for advancing lifelong AI memory.
Memory at the human lifelong scale is gaining relevance with burgeoning interest in AI that can do economically valuable human-level work and in personal AI assistants. Long-term memory remains a key challenge in the quest for AGI.
Approaches to represent and use memory remain fragmentary: retrieval-augmented generation, agent-managed tools and notes, long-context, recurrent, and state-space architectures, context compression and latent context, post-deployment and test-time training and fine-tuning, hypernetworks, and beyond.
There are few datasets and benchmarks for human-like memory at the lifelong scale, which must test deep, realistic, grounded experience and evolving environments and needs across years, well beyond small contexts and synthetic datasets.
We will publish a public synthesis of the debate and closing panel as a research agenda recording points of agreement and disagreement; a draft benchmark specification covering data units, tasks, metrics, and costs; and open questions. Contributions will be invited from panelists, debaters, and authors of relevant contributed work.
All eight speakers confirmed.
8:00 – 8:10
10 min
Opening remarks and framing
8:10 – 9:50
100 min
Morning invited session — 4 × (20-minute talk + 5-minute Q&A)
9:50 – 10:20
30 min
Coffee break and informal discussion
10:20 – 10:35
15 min
Contributed spotlight — 12-minute talk + 3-minute Q&A
10:35 – 11:35
60 min
Contributed talks — 5 × (10-minute presentation + 2-minute Q&A)
11:35 – 12:35
60 min
Debate: general-purpose AI agents versus specialized memory algorithms
12:35 – 1:50
75 min
Lunch and informal exchange
1:50 – 3:30
100 min
Afternoon invited session — 4 × (20-minute talk + 5-minute Q&A)
3:30 – 4:00
30 min
Coffee break and informal discussion
4:00 – 5:30
90 min
Poster session
5:30 – 6:00
30 min
Moderated panel to synthesize workshop topics
Tentative schedule, subject to final conference scheduling.
Can general-purpose AI agents using today's tools, including computer use, solve lifelong memory tasks without further specialized memory algorithms?
What are the biggest challenges when extending memory from individual tasks to years of experience? What experiment would distinguish competing approaches to each challenge?
We will collect audience questions and benchmark proposals before and during the workshop. The closing panel will synthesize a benchmark proposal and identify unresolved disagreements for the post-workshop article.
Alongside full papers, we are establishing a dedicated short-paper track for 3–5 page small, self-contained results, replications, reanalyses, and negative results. This track supports late-breaking work and participation by new, under-represented, or under-resourced authors.
Submissions due
February 1, 2027
Notifications
February 26, 2027
Paper submissions and reviews must be primarily human-authored. The Call for Papers requires authors to attest that submissions are not already published in a peer-reviewed venue. Workshop submissions are non-archival.
