Lesur.ai
All posts

We built the memory our agents were missing. Hivemind 1.6.3 scores 93.90% on LoCoMo.

Guillaume Lesur 12 min

Hivemind began with the desire to use the best model and harness for each job, without the hassle of transferring context. It’s a fully open-source, self-hostable memory that people and agents can share across tools.

Version 1.6.3 now scores 93.90% on LoCoMo, a benchmark that asks AI systems to recall facts and reason across long conversations. Here's where our result sits alongside the figures published by three other memory teams.

Published LoCoMo scores and company funding raised: Hivemind 93.90%, self-funded with no funding round; Mem0 92.5%, $24m; Hindsight 92.0%, $3.6m; MemOS 88.83%, more than $10 million with an undisclosed upper bound. Paired vertical bars on one chart; left axis shows scores zoomed to 88–95%, right axis shows funding from zero; MemOS shown as an open-ended range.

Other results are known; only those with verifiable evidence are included.

It started with the work we were already doing

About a year ago, my brother Christophe got me hooked on agentic development. I was following his early experiments with Transkryptor, then started building with agents myself, at work and on personal projects.

We now work across several agent harnesses, combining frontier models with open-source ones. We get better work from that mix. Different models bring different strengths, and we want to keep choosing the right one for the task.

The awkward part was carrying the project between them. One agent would work through a problem, make a decision, and leave the reasoning in its own conversation. The next agent needed the story again. With several agents working in parallel, we were spending too much effort keeping everyone up to date.

Christophe had already written Live Memory and Graph Memory for Cloud Temple and released them as open source. Together, we built on those foundations to create Hivemind. Through MCP, agents can read the project's shared memory and contribute what they learn, so the next person or agent has something to build on.

People and agents using different harnesses share a Hivemind project space. Documents have a dedicated ingestion path through automatic ontology into the knowledge graph.

The part we're especially excited about is automatic ontology. Hivemind derives the types of entities and relationships in its graph from the documents it ingests. The memory can take the shape of the work: experiments and hypotheses for a research group, programs and commitments for an NGO. The team doesn't have to design the whole knowledge model first.

That's what we mean by your organization's collective memory. We want teams, businesses, NGOs and communities to carry projects through with coordinated fleets of agents that can draw on what the group already knows. That knowledge stays with the organization when it changes tools or model providers, and the whole system can run on its own infrastructure.

What changed in 1.6.3

For memory to help, the right context has to reach the agent. LoCoMo gives us a practical way to test that: ten long conversations, with questions that require finding old facts, connecting them and getting the timeline right.

In 1.6.3, retrieved passages carry more of their source context, including relevant headings and introductions. Graph relationships also retain their endpoints and direction. The reader gets more of the structure that gives a fact its meaning. The release notes describe the changes.

Hivemind's three LoCoMo configurations: 86.75% with the original 1.6.2 documentary run, 92.73% reading the existing index through 1.6.3, and 93.90% after fresh ingestion with the frozen 1.6.3 candidate.

The fresh-ingestion run answered 1,446 of 1,540 questions correctly, a gain of 7.14 percentage points over our original 1.6.2 run. That is very encouraging. More useful still, the wrong answers show us where to work next in ingestion, graph extraction and context assembly.

The test follows mem0's public evaluator: ingest the conversations, retrieve memory, answer, then grade against the reference with a separate model call. We keep its prompts and scoring rules. Our harness adds tighter execution checks and saves the evidence at each step. Every question gets one answer; there's no search loop or picking the best of several attempts.

Next: a lot more to remember

For Hivemind 1.7.0, we're working on the BEAM harness, which covers histories from 100,000 to 10 million tokens. We want to see how Hivemind holds up as the corpus grows and the information inside it changes.

That means making document ingestion more reliable and increasing its parallelism, while improving the precision of both the graph's content and its automatically generated ontology. We'll track answer quality, latency, context use and cost, and keep the community posted as the results come in.

Hivemind is open source on GitHub. All the evidence behind this evaluation is available on request: configurations, ingestion records, retrieved contexts, prompts, answers, judge outputs, comparisons, usage records and integrity checks. Contact us directly on GitHub and ask for the LoCoMo 1.6.3 evaluation package.

Join our early-adopter community

We’re starting to bring together Hivemind’s early adopters. If you’d like to join, email us at bonjour@lesur.ai or open an issue on GitHub. We welcome your feedback and are listening to what you need from Hivemind.

We’re also happy to help you deploy it in your organization, for a side project, or for personal use.

Open the full test setup and results

We submitted the conversation sessions as 272 documents directly to LONG memory, in fresh, isolated storage. This bypasses SHORT notes and MID consolidation. The document hashes match those of the original documentary run. Ingestion received no questions, reference answers or annotated answer evidence, and generated answers were never written back to memory.

After ingestion, Hivemind retrieves memories for each question and Gemini answers from them. A separate call to the same model then grades the answer against the reference. Each question gets one retrieval, one answer and one judgment. The model cannot search again, and we don't generate several answers and pick the best.

We ran this evaluation internally using the public mem0 memory-benchmarks evaluator, pinned to commit 4b61c5d31b9c668a12b4f5e78064248a02c82d2b. We kept its LoCoMo prompts and scoring rules, including the handling of partial answers and dates, and excluded the 446 adversarial questions. The score is the percentage of answers judged correct; the original LoCoMo research used lexical F1.

The test ran on October 18, 2026, using a frozen 1.6.3 release candidate at commit 11ca440f3f2b76d1af92c79c8dfff7d6f2df316b. Both container image digests are recorded in the evidence package. The public release followed on October 19.

ComponentConfiguration used in this run
ProductFrozen 1.6.3 candidate, exact commit and both image digests recorded
Extractionclaude-haiku-5-5; temperature and effort omitted
OntologyNative ontology=auto, built separately for each conversation
EmbeddingsOpenAI text-embedding-3-large, 3,072 dimensions
Native processingExtraction slices capped at 25,000 characters; chunk size 500, overlap 50
Retrieval10 entities and 8 text chunks per question; native score cutoff retained
Reader and judgegemini-3.8-flash, reasoning effort medium, separate calls
Memory budget50,000 tokens, counted with cl100k_base, tiktoken==0.12.0
Actual memory contextMean 23,825.03 tokens; maximum 50,000
Temporal-occurrence extractionDisabled throughout ingestion and evaluation
ExecutionTen conversations in parallel; one campaign and one answer per question

We use mem0's evaluator, with tighter controls around how the run is executed and recorded. Before ingestion, we froze source versions and images, datasets, model settings, prompts and retrieval limits. We reconciled every ingestion job with the document and vector indexes, including rejected documents.

Before generating each answer, we checked where its retrieved context came from, its status and the absence of disabled temporal annotations. We saved the rendered context and prompt hashes, along with the answer, raw judge output and any truncation record. Successful operations were never repeated to seek a better score; technical failures followed a retry policy set in advance. Missing cases stayed in the planned question count, with rejected documents, recovered errors and unknown usage recorded separately.

In this run, 271 of the 272 documents were accepted and indexed. One extraction returned invalid model output; we did not resubmit it, and all 1,540 questions stayed in the evaluation. The harness cut 46 retrieved contexts at the 50,000-token ceiling. Token counts include the metadata sent alongside the passages. Retries recovered ten transient Gemini HTTP errors, leaving no answer or judgment missing.

Hivemind's recorded ingestion time was about 24 minutes 22 seconds. Model and embedding usage gives an estimated cost of $1.78, including ontology preparation and corrections. The reader and judge add about $39.09 at uncached rates. These estimates cover recorded usage, excluding infrastructure and taxes.

The three configurations

ConditionCorrect / evaluatedAccuracy
Original 1.6.2 documentary run1,336 / 1,54086.75%
Existing index with the 1.6.3 legacy-context reading path1,428 / 1,54092.73%
Fresh ingestion with the frozen 1.6.3 candidate1,446 / 1,54093.90%

Against the original 1.6.2 run, 132 previously failing answers passed and 22 previously passing answers failed. Against 1.6.3 reading the existing index, 39 improved and 21 regressed, for a net gain of 18 correct answers. This comparison covers the ingestion and retrieval changes together.

Results by question type

CategoryCorrect / evaluatedAccuracy
Single-hop808 / 84196.08%
Multi-hop266 / 28294.33%
Temporal reasoning299 / 32193.15%
Open-domain73 / 9676.04%
Overall1,446 / 1,54093.90%

All questions received an answer and a verdict: 1,446 passed and 94 failed. Each question counts equally toward the overall score.

Comparison with other long memory projects
ProductCompanyCountryPublished LoCoMoReported funding
Hivemind 1.6.3Lesur.AIFrance93.90%Bootstrapped; no funding round
Mem0Mem0US92.5%$24m, seed + Series A
HindsightVectorize AIUS92.0%$3.6m seed
MemOSMemTensorChina88.83%Tens of millions of US$

These projects explore different approaches to long-term memory. The notes below describe their published LoCoMo evaluations and methodology.

  • MemOS — 88.83%. MemTensor publishes this result for its cloud service in OmniMemEval: 1,540 questions, requested reader gpt-4.1-mini-2025-04-14, judge gpt-4o-mini-2024-07-18, and 5,400 mean answer-stage context tokens, including the prompt. The result tables, adapters and evaluation code are public. Results and configuration. Only MemOS's own score is used here.
  • Hivemind — 93.90%. Our internal evaluation of the frozen 1.6.3 candidate uses Haiku extraction and Gemini 3.8 Flash medium for the reader and judge. Mean memory context: 23,825 tokens, capped at 50,000. Full setup above; evidence available through GitHub.
  • Mem0 — 92.5%. The managed platform reports 1,425/1,540 at top-200, with 6,956 mean context tokens and one retrieval followed by one answer. The harness and per-question results are public; the result includes managed-platform optimizations absent from the OSS SDK. Results, pipeline description.
  • Hindsight — 92.0%. Vectorize published this result for v0.4.19 in AMB single-query mode on March 23, 2026. The harness, prompts and result explorer are public. Publication, harness.
  • Zep — 94.7% reported. Zep’s research page describes a LoCoMo evaluation with GPT-5.4 as reader and judge. Its methodology is public; we have not verified the underlying per-question outputs. Research and methodology.

Funding is company-level context. MemTensor's July 2026 financing announcement was reported as bringing cumulative funding to hundreds of millions of RMB; no precise total was disclosed. It followed a near-RMB100m angel round, and a further Pre-A+ round was announced in September without an amount. The table expresses this as tens of millions of US dollars, an approximate conversion rather than a disclosed dollar total. At the October 23, 2026 USD/CNY closing rate of RMB6.6703 per US dollar, RMB100m is about US$15m. Mem0's $24m covers seed and Series A together; Vectorize's $3.6m is its announced seed round. Lesur.AI is self-funded, with no funding round, as confirmed by its founder.

Country refers to the company's base: MemTensor in Shanghai, China, Lesur.AI in France, Mem0 in the US, and Vectorize in the US.