MemTensor wants forgetting to be a governed choice

A speaker addresses the OpenMem community launch co-hosted by MemTensor during WAIC 2025 in Shanghai, where researchers and operators gathered around long-running agent memory.[8]

The most consequential fact an AI agent remembers may be the one its user later asks it to forget.

A favorite hotel is harmless until it becomes a stale constraint on every itinerary. A shipping address is useful until a move. A customer-service resolution should inform the next call, but not leak into another customer’s account. Once an agent turns conversations into summaries, tool traces, reusable skills, or shared knowledge, correcting the original sentence may no longer correct its descendants.

That is the sharper way to read MemTensor, the Shanghai company behind the open-source MemOS project. On July 20, 2026, Chinese financial reporting said MemTensor had completed a roughly RMB100 million Pre-A round to fund long-term-intelligence infrastructure and “memory-native” model research.[6][9] Huawei Hubble, Honor’s strategic-investment arm, SenseTime Guoxiang, Shenzhen Capital Group, and MSA Capital were among the investors.[6]

As of 2026-07-29T21:38:16Z UTC, the company’s public stack has grown well beyond a paper proposal. MemOS 2.0 offers hosted service, a self-hosted server, cloud-connected agent plug-ins, and a local plug-in. Persistent memory is becoming a product layer with its own storage choices, identity scopes, correction paths, and deployment politics. MemTensor’s opportunity—and its burden—is to make forgetting as deliberate as recall.

For this dossier, forgetting does not mean shredding every byte. It means applying a declared policy: update current truth, revoke future use, place a logical tombstone, preserve a restricted audit record, or physically erase a source and its derivatives where required. Those operations should not be collapsed into one delete button.

The company arrived before the category settled

Founder and CEO Xiong Feiyu established MemTensor in Shanghai in November 2024 with a team from the Large Model Center he had led at the Shanghai Algorithm Innovation Research Institute. The company says it was incubated by that institute and names mathematician E Weinan as its chief scientific adviser.[1][9] Before the new Pre-A, it had raised a near-RMB100 million angel round from Futeng Capital, Suanfeng Information, CICC Capital, and others.[9]

MemTensor’s research thesis is more ambitious than “put conversation chunks in a vector database.” Its 2025 MemOS paper divides model memory into three substrates: plaintext memory such as documents and structured facts; activation memory such as KV caches and hidden states; and parameter memory encoded in weights or adapters.[3] A common container called a MemCube is supposed to carry content alongside provenance, permissions, version history, and usage signals.

The paper’s operating-system analogy is doing real conceptual work. A conventional OS does not merely store bytes; it decides which process can see them, when they move, and how their lifecycle is recorded. MemOS proposes analogous components for memory creation, activation, fusion, archiving, freezing, and disposal. Frequently used plaintext could be promoted into a faster activation form; stable knowledge might eventually be distilled into parameter modules; cold or obsolete knowledge could move in the other direction.[3]

That broad architecture establishes a useful standard against which to read the product. At the repository snapshot pinned to commit 92653cd, the most concrete, deployable surface is concentrated around plaintext, graphs, vectors, tool traces, and agent skills. The repository exposes one API family for adding, retrieving, editing, and deleting memory; multiple “cubes” for isolation and controlled sharing; and feedback paths for correcting or replacing stored material.[2] The official Chinese documentation still describes the wider plaintext–activation–parameter lifecycle, but the everyday installation path is more pragmatic than the research horizon.[4]

That gap between blueprint and product is revealing. MemTensor began with a theory of heterogeneous model memory and has entered the market through the part operators can inspect today: a service that sits between agents, models, and storage.

One product, three trust models

The current packaging divides MemOS into three materially different trust arrangements.[2]

The hosted lane gives an application an API key and keeps its data in MemOS Cloud. The same service can sit behind an agent plug-in that recalls context before a run and writes new conversation material after the run. This is the lowest-operations route, but it asks a customer to place durable behavioral history in a third party’s control plane.

The self-hosted lane runs a MemOS service with Neo4j and Qdrant on infrastructure the customer operates. That shifts database custody and network policy back to the deploying organization, while leaving it responsible for upgrades, backups, identity mapping, and the language or embedding models used during ingestion.

The local lane uses SQLite and hybrid full-text/vector retrieval inside a plug-in for OpenClaw or Hermes. Beyond offline speed, local storage changes who can inspect the memory and who must be trusted to retain it. It also changes the failure mode: a laptop migration, a corrupted database, or two agents sharing the wrong local namespace can matter more than cloud uptime.

Calling all three “persistent memory” hides these differences. A preference copied from local storage into a shared cloud cube is no longer governed by the local promise. A memory exported from self-hosting into a plug-in may lose an access rule if the formats do not carry the same identity semantics. Portability therefore requires more than moving text and embeddings; it requires moving provenance, ownership, retention state, and revocation intent with them.

MemOS’s public design recognizes much of this problem. Its paper describes version chains, namespaces, access logs, permission-aware scheduling, audit exports, and provenance attached to memory objects.[3] Its current product materials emphasize private and shared cubes, correction, and agent-level isolation.[2][4] What the public material does not yet prove is the lineage-aware result: when a source is corrected or revoked, dependent summaries, graph edges, caches, shared copies, and learned skills are recomputed, quarantined, tombstoned, preserved under restricted audit controls, or erased according to an explicit policy.

An operating system earns its name on bad memory

Recall benchmarks reward the system for finding the right old fact. Production systems also need to survive the wrong one.

Consider an assistant that records “send invoices to my personal address,” later learns a corporate address, and has already crystallized the old instruction into a reusable invoicing skill. Retrieval quality is only the first step. A governable memory system should show where the instruction came from, which artifacts depend on it, which agents can read it, and when it changed. It should distinguish updating the current address from revoking the old one’s future use; a logical tombstone from physical erasure; and a permitted, access-controlled audit record from an active memory available to the agent. Some descendants may remain valid. Others should be recomputed, quarantined, revoked, or erased.

Those distinctions become sharper as memory moves from plaintext into opaque forms. A sentence in a graph can be inspected. A promoted KV representation or a parameter adapter is harder to attribute and unwind. The MemOS paper treats cross-type conversion as a central ambition, including the possibility of distilling stable knowledge into parameter modules.[3] That makes rollback, lineage, and consent more—not less—important. A system should not promote a convenient error into a durable behavior faster than a user can contest it.

The same issue applies to shared memory. MemOS presents multi-cube composition as a way for users, projects, and agents to reuse knowledge.[2] Enterprise buyers need to know whether two agents can share one approved fact without also sharing the person, conversation, or credential that produced it. A reliable boundary makes the memory layer look like durable infrastructure; a porous one turns persistence into a larger incident surface.

The scorecard is promising—and vendor-shaped

MemTensor has made its evaluation surface more inspectable than a slide of isolated percentages. The 2025 paper reports experiments with a common GPT-4o-mini backbone on an 80 GB H800 GPU. In that setup, the authors’ MemOS-1031 configuration recorded an overall 75.80 LLM-judge score on LoCoMo and 77.8 on LongMemEval, with task-level results reported for multi-hop, temporal, preference, and knowledge-update cases.[3]

The pinned July 2026 OmniMemEval snapshot reports 88.83 on LoCoMo and 89.20 on LongMemEval for the MemOS cloud lane. Its common user-memory setup uses gpt-4.1-mini-2025-04-14 for memory-service settings and answers, with gpt-4o-mini-2024-07-18 as judge. The LoCoMo run excludes category 5 adversarial questions and covers 1,540 questions.[5] That is a useful within-harness result, not evidence that the same numbers transfer to self-hosted or local MemOS.

The benchmark most directly related to this dossier’s thesis is PersonaMem v2. Its 1,048 ask-to-forget cases test whether a system answers a preference question correctly after the user has asked it to forget something. MemOS cloud scored 57.82 by rule-matching accuracy, the highest reproduced row in that snapshot; it still missed the gold option in 42.18% of cases.[5] HaluMem offers a second behavioral signal: MemOS scored 55.00 on its 180 dynamic-update questions.[5] Neither result is a storage audit. They do not prove physical erasure, tombstone propagation, or cleanup of every derived artifact.

A separate agent track evaluates the local MemOS plug-in 2.0.8 on OpenClaw 2026.5.7, using qwen3.6-flash without thinking for the agent and with thinking for the judge. Its protocol runs cleanup, training, settling, backup, restore, and three independent executions per task across five domains.[5] The MemOS README says average completion rose from 36.63% to 50.87%.[2] Yet the five published MemOS rows—23.85, 61.00, 38.46, 64.96, and 62.07—have a simple mean of 50.07%, while the baseline rows do average 36.63%.[5] That 0.80-point inconsistency should be resolved before the headline is quoted as settled.

The boundary still matters. MemTensor authors the system, maintains the harness, selects configurations, and publishes the headline table. A July 27 market review illustrates the comparison hazard rather than resolving it: it places an older MemOS result beside a later Mem0 vendor claim, despite different versions and setups.[7] Decision-grade evidence would independently reproduce version-pinned backends while controlling ingestion budgets, retrieval depth, foundation and judge models, hardware, and product lane. It would add lifecycle audits for correction, revocation, erasure, namespace leakage, and export fidelity.

Open source meets a strategic hardware bet

The Apache-2.0 repository gives MemTensor a distribution channel, while the hosted API and self-hosted package suggest managed-service and enterprise deployment as monetization paths.[2] The local plug-in offers a third route onto devices. A 36Kr/Investment Community report says the company had signed contracts worth tens of millions of yuan by its angel round, but it does not disclose recognized revenue or how those contracts divide among cloud, enterprise, and research work.[9] ZenGen puts the new round’s valuation at roughly RMB500 million and says revenue remains undisclosed; neither figure appears in the company’s official site or Jiemian’s round report.[1][6][7]

Competition is already visible in MemTensor’s own harness: Mem0, Zep/Graphiti, Supermemory, Letta, Hindsight, and other managed or self-hosted systems occupy adjacent territory.[5] A third-party Shanghai market review contrasts MemTensor’s open infrastructure approach with Red Bear’s more closed enterprise positioning.[7] Recall accuracy alone is unlikely to hold that field apart; provenance, portability, lifecycle controls, and reliable operation across deployment lanes are more durable differentiators.

The investor list adds a specifically Chinese industry signal. Huawei Hubble and Honor are tied to hardware and device ecosystems; SenseTime brings a large AI-platform lineage; Shenzhen Capital adds state-linked venture depth.[6] An independent review of the round interprets that mix as a wager that memory may need to live close to phones and computers rather than remain an add-on to a remote model API.[7] That is an inference, not a disclosed product roadmap, but it fits MemTensor’s local plug-in and its research language about cross-platform migration.

On-device memory could give MemTensor a reason to exist even as foundation models gain longer context windows. More context lets a model read more during one run. It does not by itself provide durable identity, permissions, audit trails, correction, or a portable record that can move between models. Conversely, if model vendors and operating-system owners absorb those functions into their own platforms, an independent memory layer risks becoming a branded retrieval wrapper.

The company’s next proof is therefore organizational as much as technical. Can one memory preserve its meaning as it moves from cloud to local storage, from one agent to another, and from plaintext into a faster representation? Can a user inspect, correct, export, and erase it without understanding the internal graph? Can an enterprise verify that two tenants never inherit each other’s history? And can independent evaluators reproduce the gains after all versions and budgets are fixed?

MemTensor has raised capital to pursue the expansive answer: memory as a governed resource that outlives any one model session. The narrower store-and-search market already includes Mem0, Zep/Graphiti, Supermemory, and others.[5] The distance between those answers will be measured less by how warmly an agent says “I remember” than by how precisely the system responds when its user says, “That is no longer true.”

cronfeed.work