Prinevo Memory: A Context Graph for Long-Term Agent Memory

Long-term memory is what separates an agent that starts every conversation from zero from one that can retrieve prior facts, reason across sessions, account for knowledge updates, and abstain when the evidence is missing. Building that reliably requires more than storing chat summaries or ranking text chunks.

Prinevo Memory is a domain-pluggable context graph built around structured capture, time-aware knowledge, and retrieval strategies matched to the question being answered. We evaluated the system on two independent public benchmarks: LoCoMo and LongMemEval.

Memory needs structure, not only storage.

A flat memory store can retrieve a relevant sentence, but it struggles when the answer depends on several distinct facts, a change over time, or an exhaustive list. Prinevo captures memory as typed subject-predicate-object claims linked to resolved entities. That structure makes grouping, multi-hop traversal, temporal comparison, and cardinality-aware retrieval possible.

The memory engine is shared across domains. A swappable ontology profile defines the entity types, relationship types, and dimensions that matter for a domain, while the capture, storage, retrieval, and answering machinery remains the same. New entities and relations resolve against the existing catalog or are minted from real usage when genuinely new.

How the context graph works.

1. Typed-edge capture: turns session or thread text into structured claims with entity references instead of relying on prose summaries.
2. Entity resolution: merges a mention only when confidence and separation from the next candidate are sufficient; otherwise it mints or defers rather than guessing.
3. Bi-temporal supersession: records when a claim was observed and when it was true, retiring outdated claims without erasing the history.
4. Hybrid retrieval: combines relevance, recency, and entity-anchored candidate lanes with reciprocal-rank fusion.
5. Question-aware answering: classifies cardinality, temporal axis, and answer mode before selecting a point lookup, exhaustive gather, comparison, or aggregate strategy.
6. Abstain recovery: widens retrieval after a first-pass evidence miss before the system gives up.

Every score needs its methodology.

Benchmark percentages are sensitive to the judge, prompt, sample size, corpus, and model configuration. The paper therefore reports a published-style headline score and an alternate strict-judge score for each benchmark. Both judges score the same final answer set.

Both benchmark runs used Claude Haiku 4.5 for capture and answering, with Claude Sonnet 5 for judging. Prinevo uses its own stricter reasoning-then-verdict rubric during development, even when a different published rubric is used for external comparability.

Where the remaining errors live.

The category results show that memory quality is not one problem. On LoCoMo, adversarial abstention and single-hop questions are strongest, while multi-hop and open-domain questions remain harder. On LongMemEval, temporal reasoning reaches 90.98%, while single-session preference questions are the weakest category under the benchmark's own judge.

An earlier full LoCoMo error analysis found that only 2% of wrong answers were true extraction misses. In 63% of wrong answers, the gold fact was already present in the retrieved evidence but the answer was assembled incorrectly. Another 16% had the fact in evidence but abstained, and 19% lacked the required evidence, primarily on adversarial or inference questions. The largest remaining opportunity is therefore selecting and assembling available evidence correctly, not simply capturing more facts.

An agent's answer quality is bounded by the evidence and structure it reasons over, not only by the model generating the final response.

Why domain-pluggable memory matters.

The evaluated profile focuses on conversational memory, but the same engine also supports an engineering and repository-analysis profile with a different ontology. Product behavior, owners, repositories, contracts, incidents, decisions, and rollout history can become typed, reusable context without rebuilding the underlying memory architecture for each domain.

Early engineering Q&A testing has shown roughly a 54-percentage-point uplift with memory versus without it. That domain result is not yet a published benchmark, so it is not presented as directly comparable to the LoCoMo or LongMemEval results. The next step is to publish the same judge, sample, and methodology detail for engineering context.

What these results do not claim.

The evaluation is not a same-judge, same-corpus comparison against competing memory systems. Published numbers across the field use different judges, prompts, sample cuts, and sometimes different metrics, so they should not be treated as a direct ranking without a controlled reproduction.

LoCoMo also has documented ground-truth inconsistencies that impose a small ceiling unrelated to the memory system itself. Prinevo does not tune toward known incorrect labels. The continuous multi-agent case, where several agents update and reuse the same context graph throughout delivery, is also still being measured and is not implied by the conversational benchmark results.

What we learned.

Better retrieval cannot recover structure that was never captured. Typed claims, entity resolution, time-aware supersession, and question-aware answer strategies have to work as one system. A raw storage layer does not decide what is worth remembering, whether two mentions refer to the same entity, or how an outdated fact relates to the current one.

The durable claim is not one percentage in isolation. It is that each result can be checked against a named judge, prompt source, corpus, and sample size. That transparency makes the benchmark useful for improving the system instead of using it only as a marketing number.

Related reading.

• Pillars of an Autonomous Agentic System
• Building a Software Factory: From Prompts to Compounding Systems