Mem0’s Memory Build Cost More Than Full Context, Study Finds

Building the memory store of Mem0, an agent-memory framework, cost more than answering every question with the whole conversation in the prompt, with each of the three AI models tested, according to a study by Zedong Peng of Shanghai Jiao Tong University posted on arXiv on 21 October 2026. With Claude Opus 5.5, Mem0’s build alone cost $25.88, against $17.84 for an entire full-context run, the paper reports.

Agent memory, in the paper’s sense, is a system such as Mem0 or Zep that decides what to keep from a conversation before any question arrives: the model extracts facts or builds a graph, and each question then sees a few retrieved pieces. Full context means sending the entire conversation history with every question. The third option tested is a coding agent, such as Pi, Codex or Claude Code, that keeps the raw history and searches it with its own tools.

With Claude Opus 5.5, the Pi coding agent scored 99% for $6.95 a run, under half the cost of full context, the paper says. Peng names the gap: “Paying to build memory that a stronger model does not need is a memory tax.” He also states the study’s size up front: one conversation and one run per setup.

Business Pill · AGENT MEMORY

A short Business Engineer explainer on agent memory: how an AI assistant keeps what it learns between conversations. It teaches the general idea only and says nothing about the study in this piece.

What the Paper Found

• The build bill. Mem0’s build alone cost more than full context’s whole run with the same model: $0.56 against $0.43 with GPT-4o-mini, $0.45 against $0.29 with GPT-6 Luna and $25.88 against $17.84 with Claude Opus 5.5, the paper reports. Mem0’s questions are cheaper, so the paper says the build would pay for itself after 372, 545 and 596 questions on the same history; the test asks 152.
• Where memory helped. With the weakest of the three models, GPT-4o-mini, Mem0 scored highest: 78%, against 74% for both full context and BM25, a standard keyword search, according to the paper.
• Where it did not. With Claude Opus 5.5, Mem0 and full context both scored 95% and Zep 87%, while the three coding agents scored 99% (Pi), 98% (Codex) and 97% (Claude Code), the paper reports.
• A ceiling set by the writer. On the Zep memory graph that GPT-4o-mini wrote, the three models scored 43%, 43% and 45% as readers; on the graph Claude Opus 5.5 wrote, 66%, 71% and 87%. The paper puts it this way: “A stronger reader does not recover what the writer left out.”

Source: Zedong Peng, ‘MemoryTax: The Bitter Lesson of Agent Memory’, arXiv 2610.29823 v1, 21 October 2026. One conversation, one run per setup. All figures are the paper’s.

What Was Posted on X

Elvis Saravia, who posts as omarsar0, shared the MemoryTax paper on 22 October and called it great insights for agent builders. In his summary, Mem0 beats plain full context only with GPT-4o-mini on the 419-turn conversation, and building Mem0’s memory costs more than an entire full-context rollout. His closing advice, that the study is a cheap experiment to repeat before adding a memory layer, is his framing.

How the Test Worked

The study uses one conversation, conv-26, from LoCoMo, a public benchmark of long chat histories: 419 turns in 19 sessions and 152 questions, the paper says. The questions cover single facts, facts that must be combined, dates and open-ended recall. The whole history fits in every model’s context window, so sending all of it is always possible.

Peng ran 21 setups, seven methods with three models each, and ran each setup once. A GPT-4o-mini model graded the answers. Costs are list prices for every model call, with the grader excluded, the paper says. Pi read about 10,000 input tokens a question, 7,900 of them cached, against 28,900 for full context.

The One Miss and the Photo Caveat

Pi’s single miss with Claude Opus 5.5 was a faulty answer key, according to the paper. The benchmark says Caroline shared an abstract painting, while in the conversation it is Melanie who shares it; Pi said so and was marked wrong. The paper says full context, Codex and Claude Code flagged the same mistake.

The agents could also read photo captions that the memory methods were not given, and 64 of the 152 questions had evidence in a turn with a photo. On the 88 questions without photo evidence, Pi scored 100% and full context, Mem0, Codex and Claude Code 99% each, the paper reports. Peng writes: “Whether searching the record answers more than reading all of it, this study does not settle.”

What Elvis Saravia Posted

Elvis Saravia, who posts on X as omarsar0, shared the paper on 22 October and called it “Great insights for agent builders; bookmark it.” In his summary, “Building Mem0’s memory costs more than an entire full-context rollout.” His closing advice is his own framing: the study is “a cheap experiment to repeat on your own agent before adding a memory layer.”

Connecting the Coverage

Our LangChain memory piece reported on 18 October that LangChain chief executive Harrison Chase said on the 20VC podcast, “I think memory is super, super early.” Our memory-store guide covered on 20 October an essay by Alex Wang arguing that a linear layer, an MLP and self-attention are three kinds of memory store.

Our Nadella harness piece reported that Satya Nadella’s 21 October shareholder letter describes a harness, “including context, memory, and action space”, that is separate from the model. Our peer-review piece reported on 23 October that three former journal editors preferred Claude Opus 4.8’s reviews over paid human ones in 26 of 30 blind picks. What’s new: this paper puts list-price figures on building a memory layer and sets them against letting the model read or search the raw history.

The Structural Read

As we read it, the memory tax falls on whoever pays for the build: in the paper’s tests, Mem0’s build with Claude Opus 5.5 cost $25.88 before any question was answered, against $17.84 for a whole full-context run.

As we read it, the comparison that matters for memory-layer companies is against better search tools, not against no memory: the paper says a strong model may need “exact, semantic and vector search over the raw record, driven by the model itself”.

As we read it, the paper leaves designed memory one clear case: it says designed memory “may survive as a cost trade-off when many questions share one history”.

Three Implications

FOR AGENT BUILDERS The paper’s test used one conversation of 152 questions, ran each setup once and gives the cost of every run at list prices.

FOR MEMORY-LAYER VENDORS In the paper’s results, Mem0 led only with GPT-4o-mini, 78% against 74% for full context, and its build cost more than a full-context run with all three models.

FOR BUYERS OF AI AGENTS The paper reports that the same model cost $6.95 a run in Pi, $25.41 in Codex and $28.41 in Claude Code, and points to a separate study, HarnessTax, on why the costs differ.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework states: “Orchestration, memory, tool use, and workflow execution have become the new sources of competitive advantage.”

As we read it, the MemoryTax paper sets two of those against each other, designed memory and a model using search tools, and in its one conversation the tool-using agents scored higher with Claude Opus 5.5. The framework was published in July 2026 and does not discuss this paper.

What Is Not Established

The paper lists its own limits. It tests one of LoCoMo’s ten conversations with one run per setup, and says differences of a few questions may not hold in a second run. A GPT-4o-mini judge with a lenient rubric did the grading, and at least one answer key is wrong.

The agents saw photo captions the memory methods did not get. The history fits every model’s context window, and the paper says results may differ on other benchmarks and on histories that do not fit. Mem0 and Zep ran with default settings, and costs are list prices that depend on each provider’s caching. The paper is a version 1 arXiv preprint by a single author. We did not contact the author or rerun the study. Nothing here is investment advice.

The Bottom Line

In this one-conversation study, building Mem0’s memory cost more than a whole full-context run with each model tested, and with Claude Opus 5.5 a coding agent that searched the raw history scored 99% for under half the full-context cost. The author says the results suggest a strong model needs less designed memory and better tools for searching the raw record, and lists the limits that keep it a small study.