Call for Papers: PALM at NeurIPS 2026

AI is losing its memory problem and gaining a memory field.

For most of the last decade, the frontier question was reasoning inside a single context. How much can a model hold at once, and how well can it use it. That question is being replaced by a harder one. Not what a system can process in a moment, but what it should carry across them.

Conversational assistants are now expected to remember users between sessions. Agents are expected to hold task histories, tool traces, and shared state across long horizons. Embodied systems are expected to build spatial memory of places they have already been. This is a shift away from single-context interactions and toward persistent memory layers. Systems that encode, retrieve, update, and sometimes forget.

That shift raises questions the field has not yet answered together. How should an AI system decide what to remember. What should it forget, and when. How do you evaluate memory beyond short-context recall. How do you keep persistent memory controllable, transparent, and safe when it can be poisoned, leaked, or quietly manipulated over time.

PALM exists to build a shared research agenda around exactly these questions.

About the workshop

PALM is a workshop at NeurIPS 2026 in Paris, focused on safe, personalized, and long-term memory for AI agents and conversational assistants. It brings together researchers across memory architectures, personalization, multimodal and embodied memory, agentic and multi-agent memory, cognitive and neuroscience-inspired memory, evaluation, and safety.

The organizing committee spans CISPA Helmholtz Center, KAIST, the ELLIS Institute Tübingen and MPI-IS, Google DeepMind, the University of Bristol, and Memories.ai. Our co-founder Shawn Shen is a co-organizer. We are helping to convene this workshop because the memory layer is the problem we work on every day, and we think it deserves a venue of its own.

Why we care about this one

Memories.ai builds long-term memory infrastructure for the physical world. This video, sensor streams, and the operational environments where machines actually work. The hardest version of the memory problem is not textual. The physical world does not arrive as clean tokens. It arrives as long, redundant, high-bandwidth visual data that has to be encoded, retrieved, and reasoned over across time.

That is why one of the workshop's seven topic areas matters to us in particular: multimodal, visual, video, and embodied memory. Long-term video memory, visual retrieval for agents, spatial memory for embodied systems, memory for robotic and AR/VR assistants. If you work on any of these, this is a room worth being in.

But the workshop is deliberately broad, and so is the call. We want the architecture people and the neuroscience people and the safety people in the same building.

What we are looking for

We're big on research and invite submissions on long-term memory for AI agents, conversational assistants, and personalized AI systems. Work is welcome from machine learning, NLP, AI agents, HCI, cognitive science, neuroscience, privacy, security, and AI safety.

Submissions may present new architectures, benchmarks, datasets, evaluations, systems, theoretical perspectives, position papers, negative results, or interdisciplinary analyses. Negative results and position papers are genuinely welcome — a shared agenda needs disagreement and dead ends as much as it needs wins.

Topics of interest include, but are not limited to:

• Memory architectures for conversational assistants: Writing, retrieving, updating, consolidating, and forgetting across sessions; handling stale or contradictory memories.
• Memory for LLM agents and multi-agent systems: Persistent task histories, tool-use traces, shared memory, memory provenance and isolation across agents.
• Multimodal, visual, video, and embodied memory: Long-term video memory, visual retrieval for agents, spatial memory, multimodal event memory, memory for robotic and AR/VR assistants.
• Neuroscience-inspired and cognitive memory models: Complementary learning systems, episodic-to-semantic consolidation, replay, forgetting, cognitive maps.
• Benchmarking and evaluation: Long-horizon benchmarks, temporal reasoning over past events, update and deletion tests, contradiction handling, abstention under uncertainty.
• Safety, privacy, and security: M<emory poisoning, prompt injection through stored memories, sleeper memories, privacy and cross-domain leakage, long-term manipulation, alignment drift.
• User control and transparency: Inspecting, editing, deleting, and scoping memories; consent and access control; provenance; right-to-be-forgotten mechanisms.

How to submit

Tracks. Full-length papers up to 9 pages, or short papers up to 4 pages. Page limits exclude references and supplementary materials.

Format. Use the NeurIPS 2026 template. Submissions must be fully anonymized for double-blind review, including any linked code or supplementary material.

Non-archival. The workshop is non-archival. Ongoing and unpublished work is welcome, as are papers under review or recently accepted elsewhere, provided this does not breach the dual-submission or anonymity policies of the other venue. You can submit the same work to other venues concurrently or later. Accepted papers are made public; rejected submissions and reviews are not.

A note on AI assistance. AI-generated papers are not permitted. AI assistance is fine, but submissions must be primarily human-authored and reflect original thought.

Submit on –> OpenReview

Key dates

• Submission deadline: August 24, 2026
• Notification of acceptance: September 29, 2026
• Camera-ready: TBD
• Workshop: December 12 or 13, 2026, at NeurIPS 2026 in Paris (exact day TBA)

All deadlines are 11:59pm AoE (Anywhere on Earth).

If you are building the memory layer — how AI systems should remember, what they should forget, and how to make persistent memory safe — we want to read your work.

Submit by August 24. We will see you in Paris.

Workshop site: https://palm-neurips-2026.github.io/

‍Questions: [email protected]

‍