Agent memory · LongMemEval · June 2026

Building and evaluating long-term conversational memory

We built three approaches to retaining information across conversations—raw retrieval, extracted memories, and a hybrid of both—and evaluated them on 500 LongMemEval questions. The hybrid approach reached 86.1% task-averaged accuracy.

500-question benchmark 6 task types gpt-4o answer model and judge

Background

Why conversational agents use external memory

LLM-based assistants often need information from earlier conversations: user preferences, facts supplied by the assistant, or updates spread across multiple sessions. The simplest approach is to place the full conversation history in the model’s context window for every new request. As that history grows, prompts become longer, slower, and more expensive.

External memory offers another approach: store information from each session and retrieve only the context relevant to the current request. LongMemEval tests how well systems do this. Its Small split contains 500 questions over multi-session chat histories, covering user and assistant facts, preferences, knowledge updates, evidence spread across sessions, and temporal reasoning.

For each example, the memory strategy first ingests all prior conversation sessions and builds its memory store. We then submit the target test question, retrieve relevant memories, and add them to the answer prompt. The generated answer is compared with the reference answer by an LLM judge using LongMemEval’s official benchmarking scripts.

Memory strategies

Three approaches

01

Instruct

LLM-extracted memory

An extractor reads each session and the current memory store, then creates, updates, or deletes individual facts. This can consolidate changes across sessions, but details omitted during extraction are unavailable later.

sessionextract factsretrieve
02

Remis

Semantic retrieval from raw conversations

We call this strategy Remis. It splits sessions at topic shifts and indexes the resulting chunks with dense and BM25 signals. Neighboring chunks expand the retrieved context. This preserves exact details, but it can also return more irrelevant text.

sessionhybrid indexraw excerpts
03

Remis + Instruct

Raw excerpts + extracted facts

Both stores are queried, then one answer call receives two labeled sections: conversation excerpts and extracted memories. The answer model can draw from either source when responding.

raw excerpts+factsanswer

Standalone Remis is described because it supplies the raw-conversation path in Remis + Instruct. The results focus on Instruct and the combined configuration.

Evaluation overview

What the evaluation measures

Accuracy. We report task-averaged accuracy across six LongMemEval question types.

Strategies. The main comparison is between Instruct and Remis + Instruct.

Cost. Estimates reflect observed model usage at list prices.

Results

Results for the tested strategies

The figures show overall accuracy, task-level results, and comparisons with other memory systems.

Complete runs

86.1% Remis + Instruct

71.2% Instruct

Higher accuracy across all six task types

The largest differences were on preferences and assistant facts. User facts were already strong with Instruct alone.

Accuracy by task type

LongMemEval small (%)

Instruct Remis + Instruct
Each task type contributes equally to the overall metric.

Task-averaged accuracy*

LongMemEval small (%)

Our configurations Measured third-party Published only
* Spring 2026 snapshot; competitor systems may have changed since evaluation.

Comparison systems

Remis + Instruct led the comparison

Remis + Instruct reached 86.1%, followed closely by Oracle at 86.0% and Mastra OM at 84.2%. Orange rows are systems we measured; gray rows are published reference values. The chart provides broad context rather than a controlled head-to-head ranking.

Why Remis + Instruct works

Extracted memories can consolidate information spread across sessions and represent changes over time. Raw chunks preserve quotes, names, dates, and numbers that an extractor may omit. Remis + Instruct gives the answer model access to both forms of context.

A July 2026 preprint, Beyond Memory Leaderboards, provides a useful comparison. In its scientific-retrieval evaluation, adding lexical BM25 to dense semantic retrieval was the largest intervention measured. Under a matched retrieval budget, the three resulting sparse–dense hybrids were effectively tied at the top of its PTr benchmark.

That design maps directly to the Remis path: Remis combines dense semantic search with BM25 over raw conversation chunks, while Instruct contributes separately extracted facts. Remis + Instruct therefore pairs sparse–dense raw-context retrieval with structured memory. The paper studies scientific documents rather than conversations, so it supports the retrieval pattern rather than validating our LongMemEval result.

Cost

Cost in context

We include systems for which our evaluation records contain both an accuracy result and enough token data to estimate visible LLM spend.

Accuracy and cost

Highest accuracy at about one-fifth of Mastra’s modeled cost

Among complete Small-split runs with fully observed model-token spend, Remis + Instruct scored 1.9 points above Mastra OM at $33.25 versus $171.65 per million conversation tokens. Instruct was cheaper but 14.9 points less accurate. Gray lower-bound points omit extraction or managed-ingestion spend, so their apparent cost advantage is incomplete.

Measured systems

Upper-left is better; cost axis is logarithmic

Our configurations Complete third-party Partial or lower bound
Modeled visible LLM spend per 1M conversation tokens. All points come from our evaluation runs; arrows mark lower bounds. Infrastructure and hosting are excluded.

Reasoning-effort study

How reasoning effort affects extraction accuracy

We varied reasoning effort within each extraction model and measured the resulting accuracy and cost.

Reasoning effort, extraction accuracy, and cost

Higher accuracy and lower cost are better

Upper-left is better
Accuracy versus cost for gpt-5-nano, gpt-5-mini, gpt-5, and gpt-4o as memory extraction models across reasoning-effort settings

Result

More reasoning did not always improve accuracy

More reasoning generally increased cost, but it did not guarantee better extraction. The best result came from GPT-5-mini at medium effort, the highest effort tested within that model family.

Summary

Raw evidence and extracted facts produced our strongest result

On LongMemEval Small, Remis + Instruct was the strongest configuration we measured: 86.1% task-averaged accuracy, with a higher score than Instruct in all six task types. The result supports combining hybrid retrieval over raw conversation chunks with extracted facts for long-term conversational memory.