Get Started
Topic · 3 recaps
AI Memory
Research on how AI systems retain, retrieve, update, and forget information across tasks and interactions.
Play all
...
Overview
Posts
Questions
Home
Topics
Search
Library
Questions researchers are working on
Follow a question through Rcap’s explanations and the latest papers addressing it.
Search
Can language agents maintain hidden state consistently across dialogue branches using only public conversation history?
A chat interface exposes conversation history but provides no separate channel for state that must remain hidden. When dialogue branches, the agent must preserve the same secret and answer consistently without revealing or reconstructing it from public text.
How can 70B language models fit on one GPU while preserving long-context speed and accuracy?
A 70B model must fit its weights and growing KV cache within one GPU’s limited memory. Long prompts make compression choices affect both decoding speed and model accuracy.
How can a model select the correct entity-specific parametric memory for graph queries without query-time text or semantic similarity?
Graph facts stored in separate parameter memories avoid repeating retrieved subgraphs in each prompt. However, a query’s answer may reside in one specific entity’s memory even when another entity is semantically similar, making ordinary similarity retrieval unreliable.
How can a navigation map encode heterogeneous route costs compactly while answering new goals without retraining?
A reusable map must preserve nonuniform additive edge costs without storage that grows too quickly as the environment expands. It must also support new goal queries from one learned representation rather than rebuilding or retraining for each goal.
How can aerial visual place recognition adapt across missions without catastrophic forgetting?
Aerial place-recognition models encounter substantial visual changes across successive missions even when the geographic locations remain fixed. Updating the model for new conditions can degrade recognition of environments learned earlier.
How can agent runtimes avoid context poisoning and latency from growing histories during long-horizon skill execution?
Long-running agents often append observations, actions, and intermediate reasoning to their conversation history. As that history grows, execution becomes slower and earlier context can interfere with later decisions.
How can AI agents adapt execution routes as runtime evidence invalidates their planned continuation?
Agents often commit to a multi-step execution route before intermediate outcomes are known. When those outcomes change what should happen next, stale actions or broad replanning can waste computation and discard useful progress.
How can AI memory remain inspectable and directly correctable during long-term interaction?
During long-term interaction, models may implicitly revise or create information that guides later inferences, while users cannot see or directly correct those updates. This makes it difficult to keep the system’s evolving understanding aligned with the user.
How can AI-agent memory survive model upgrades without losing retrieval or repairability?
A new model may interpret old notes differently or query existing embeddings differently, even when the memory store is unchanged. Summarization can discard evidence, while embedding changes can make stored information harder to retrieve or repair.
How can assistants remember and reason about how users sounded across long, multi-session conversations?
Transcripts preserve words but can discard emotion labels, prosody descriptors, and voice events. Assistants working across long, multi-session histories may therefore fail on questions whose answers depend on how a user spoke.
How can associative-memory capacity be compared across Hopfield and attention-like models without conflating assumptions?
Hopfield-style memories combine recurrent dynamics, energy landscapes, and pattern storage, but retrieval capacity depends on the disorder ensemble, scaling limit, and success criterion. Connections to attention and biological interpretation introduce further assumptions that can make superficially similar results incomparable.
How can asynchronous AI agents resist false claims gaining authority simply because they arrive first?
When agents exchange messages asynchronously, the first claim received can shape later answers even when it is wrong. Context and provenance may be lost across handoffs, making verification harder.
How can autonomous LLM agents detect attacks whose evidence accumulates across loop iterations?
Autonomous agents may encounter attacks whose evidence is split across multiple iterations. Safeguards that reset at each trajectory can miss this cumulative pattern.
How can autoregressive language models add recurrent cross-token memory without breaking KV caching or one-forward-per-token decoding?
Standard transformers rely on attention and cached keys and values during generation, but they do not naturally carry a separate recurrent state across tokens. Adding one can alter the decoding interface or require additional forward passes.
How can camera-controlled video generation preserve spatial consistency over long horizons despite noisy 3D memory?
Long camera trajectories require generated frames to remain aligned with the scene over time. Multi-view 3D reconstruction can place the same surfaces at inconsistent locations, making history-based geometric conditioning noisy and degrading spatial consistency.
How can coding agents maintain accurate repository context as codebases evolve without repeated discovery?
Coding agents repeatedly search, navigate, and retain repository information, while code edits can invalidate indexes and prior context. Disconnected tools and task-local histories make it costly to keep retrieved locations and context aligned with the current commit.
How can computer-use agents retain, refine, and reliably reuse procedural skills across repeated GUI tasks?
GUI interaction experience is typically transient, so procedures learned during one rollout may not remain available or reliable on later tasks. Repeated skill revisions can also produce retrieval across tasks and revision churn, making long-term behavior difficult to interpret.
How can conversational agents retrieve the right memories when users rely on implicit conversational context?
Conversational users may refer to earlier information indirectly without stating which fact they need. Systems that perform well on explicit fact questions can therefore miss memories needed to ground an appropriate response.
How can data-centric agents reuse prior intent-to-object discoveries without causing retrieval interference?
Data-centric agents often discard mappings between task intent and relevant data objects after discovery, forcing similar searches to start over. Reusing those mappings may improve retrieval, but stale or conflicting memories can interfere with later discovery.
How can embodied vision-language navigation follow instructions in unseen environments with limited data and memory?
A navigation agent must translate language instructions into actions while moving through unfamiliar environments. Maintaining enough history for each decision can require cognitive maps, accumulated frames, or external 3D tools, while next-action supervision can demand substantial training data.
Previous
1 / 4
Next