Can evolvable cognitive memory improve long-term dialogue reasoning?

Flat retrieval confuses opinions with records

The paper's starting problem is not storage capacity but epistemic status: whether a retrieved sentence is an event record, a fact, or someone's opinion about another person. In a dense vector space these can look nearly identical. A controlled probe reports that matched fact–claim pairs have an average BGE-M3 cosine similarity of 0.8231, and that flat retrieval misattributed the speaker's opinion as the requested record in 64 of 150 cases, or 42.7%. The authors call this failure Semantic Collapse and link it to memory contamination that undermines Theory-of-Mind reasoning.

The second limitation is intentionality. A fixed retrieval pipeline cannot adjust its search strategy after seeing intermediate results, which the authors argue is needed to connect evidence dispersed across sessions. CogMem is their response: a training-free architecture that converts each dialogue turn into provenance-aware graph records and then reconstructs query-relevant context through selected operators.

Claim nodes keep the speaker attached

The PEC2F schema has five node types: Person, Event, Concept, Claim, and Fact. Person and Concept nodes act as relatively stable anchors; Event nodes record time-bounded actions; Fact nodes record observed or self-reported attributes and habits. The paper is explicit that Fact is an operational record type, not a guarantee of independently verified truth.

Claim nodes carry the subjective layer. A statement such as "Jon thinks Gina is lazy" is stored as a two-hop path from the person to a Claim node and from that Claim node to its target, with content and sentiment held in the Claim. This extra hop means that even when embeddings are similar, the graph structure records whether information is a source-attributed belief or an unattributed record. Claims from different speakers are never merged, so disagreement survives reconciliation.

Consolidation compresses episodes but keeps evidence

CogMem ingests dialogue incrementally. An extraction step maps time-bounded actions to Event nodes, stable self-reported attributes to Fact nodes, and opinions about another person or concept to Claim nodes with explicit source and target fields. Every extracted node retains its source-turn identifier and timestamp, which later lets the Evidence Grounding operator recover the original utterance.

Consolidation runs offline as a graph-rewriting step. When a cluster of Event nodes linking one person to one concept reaches a threshold of three events within a window of at most 30 days, an LLM-parameterized function abstracts the pattern into a higher-level Fact node with a temporal envelope. Rather than deleting the episodes, CogMem adds SUPPORTED_BY edges from the new Fact to each source Event. The agent can therefore answer a high-level question from the consolidated Fact or follow those edges down to the original episodes when a user asks for specifics.

Claim reconciliation is narrower. Claims are reconciled only when they share a canonical source, target, and predicate; they are ordered temporally and marked CURRENT, SUPERSEDED, or CONFLICTING, with the original Claims still accessible.

Four operators replace fixed k-hop expansion

Retrieval in CogMem is a rule-based controller driven by LLM intent parsing. The operators and routing rules are deterministic, while the intent label and the judgment that evidence is sufficient come from the LLM. Anchoring always runs first, mapping the query to entry nodes with a hybrid score that combines fuzzy lexical matching and BGE-M3 embedding similarity, with lexical weight 0.3, returning up to 15 anchor nodes.

The other three operators do different jobs. Spatiotemporal traversal filters neighbors by edge type and a time window extracted from the query, with at most 15 results per hop and a maximum depth of two. Intersection computes semantic overlap across the neighborhoods of multiple entities, using 10 candidates per entity and relations such as LIKES, USES, and PARTICIPATED_IN. Evidence Grounding retrieves the source text identified by a node's stored source id, validating graph inferences against the original context. The controller stops when evidence is sufficient or after at most five post-anchoring steps.

The paper's case traces show the division of labor. A question about what two people both enjoy anchors both people and then intersects their Concept neighborhoods before verifying LIKES edges. A question about when someone moved from Sweden traverses the event timeline with a time filter and then grounds the extracted year in the source text.

Gains concentrate on multi-hop and update tasks

The evaluation covers LoCoMo and LongMemEval, two English long-term dialogue benchmarks, with two LLM backbones. The paper reports strong performance overall, and states that results are especially strong on multi-hop, temporal, and knowledge-update tasks. Ablations and the semantic-collapse probe are presented as support for complementary contributions from epistemic separation, consolidation, and agentic retrieval.

These are aggregate claims about benchmark performance rather than a single headline number, so the appropriate reading is comparative: the architecture is argued to help most where a query requires connecting dispersed evidence or tracking a changed belief. The paper does not claim that the approach eliminates retrieval error.

Remaining errors point to embeddings and time

An error analysis of 100 randomly sampled LoCoMo test cases, 25 per task category, locates the residual failures. Vector semantic drift accounts for 23% of analyzed failures overall and is highest for open-domain questions at 32%; the authors note that schema structure does not eliminate embedding-space errors in fine-grained distinctions. Incomplete search scope accounts for 15%, because strict separation of Claims, Facts, and Events forces the agent to search multiple node types and the policy does not automatically expand node types when an initial search returns nothing. Schema-granularity errors from consolidation account for 11%, as when specific qualifiers are abstracted into general categories.

Temporal queries have a distinct profile: temporal normalization errors account for 44% of errors within that category, far above the other categories. The reported cause is that the ingestion module stores the session date but does not resolve relative expressions such as "last year" against the dialogue timestamp, so the system can return the dialogue date instead of the referenced absolute date. The authors propose conditional consolidation, automatic node-type expansion, and hybrid lexical–dense retrieval as fallbacks.

What the evidence supports for practice

The practical priority this paper establishes is provenance over volume. If a long-term assistant must answer questions about who believes what, or about how a person's stated position changed, the evidence points to keeping source-attributed Claims structurally separate from Event and Fact records and to reconciling only same-source, same-target, same-predicate updates. The reported 42.7% misattribution rate in flat retrieval is the concrete reason to treat that separation as a design requirement rather than a refinement.

The same evidence sets the boundary. The evaluation covers two English benchmarks and two backbones, the semantic-collapse probe is synthetic, and the human preference study contains only 50 examples with two primary annotators and one adjudicating annotator; LongMemEval also relies on an LLM judge. CogMem should be read as making long-term memory more structured, attributable, and auditable, not as removing retrieval error. The most decision-relevant split is task type: the architecture's reported advantage is strongest where evidence must be connected across sessions or updated over time, while open-domain and temporal questions remain the places where embedding drift and date normalization still decide the answer.

Sources used in this answer