The science behind Context Lattice.

The Core Problem with Passive Retrieval

LangMem, Mem0, and conventional RAG share the same fundamental architecture: retrieve memory entries by semantic similarity to the current query, inject the results into the LLM's context window, and stop. The paper names this passive retrieval — a stateless policy that selects all memory units upfront, based solely on the query, with no adaptation during inference.

The paper documents three resulting failure modes: the retrieval strategy cannot be revised mid-reasoning; fixed similarity scores return surface-level matches that flood the context window with irrelevant noise; and static relevance functions cannot scale across unpredictable, long-horizon interactions.

The Cognitive Neuroscience Shift

The researchers root their alternative in cognitive neuroscience, which does not model memory as passive read-out of stored content. Instead, retrieval is an active and associative reconstruction process: initiated by contextual cues, propagating through intermediate representations, progressively reconstructing coherent memory experiences from partial evidence.

This shifts the engineering question from "what is semantically similar to this query?" to "what evidence do I need to reconstruct next, given what I've accumulated so far?" The first is a lookup. The second is a reasoning loop. The paper proves the two are not equivalent — and shows which wins.

The Cue–Tag–Content Graph

MRAgent organizes memory as a heterogeneous graph. Cues are fine-grained keywords — entities, attributes, contextual keywords. Contents store the actual memory items. Tags connect them, encoding associative relations between cues and content units. Three memory layers organize the graph:

Memory is populated automatically. The authors designed an ingestion pipeline that uses LLMs to process raw interaction histories and populate the graph — developers orchestrate the pipeline, not the labeling.

The Reconstruction Loop

Benchmark Results

Tested on LoCoMo and LongMemEval — benchmarks that evaluate agents on long-horizon tasks spanning dozens of sessions and hundreds of dialogue turns. Backbone models: Gemini 2.5 Flash and Claude Sonnet 4.5. MRAgent outperformed every baseline across both models and all question types.

The Semantic Similarity Ceiling

Conventional retrieval — BM25, dense retrievers, rerankers — operates on the same assumption: pre-rank documents by similarity to the query, select the top-k, and inject them into the LLM's context window. The paper calls this a single-pass ranking paradigm: relevance is determined once, before reasoning begins, and never revisited.

This works for simple factoid queries. It fails on complex, multi-step information needs where the agent cannot know what to search for until it has seen intermediate results. The paper demonstrates that even strong reranking pipelines miss relevant documents that only become discoverable through iterative exploration.

Direct Corpus Interaction

DCI replaces passive ranking with active corpus exploration. Instead of querying a vector index, the agent interacts directly with the raw corpus using general-purpose tools — issuing searches, examining specific documents, and adapting its retrieval strategy based on what it finds mid-reasoning.

The key insight: retrieval quality depends not only on reasoning ability but also on the resolution of the interface through which the model interacts with the corpus. A high-resolution interface — one that permits reformulation, targeted document inspection, and iterative refinement — unlocks retrieval accuracy that single-pass systems cannot reach regardless of embedding quality.

Multi-Benchmark Results

DCI was evaluated against BM25, dense retrievers, rerankers, and hybrid approaches across three benchmark families — each targeting a different retrieval difficulty level.

Why Agents Beat Embeddings

The paper identifies three capabilities that DCI agents have and single-pass retrieval systems lack: