Four papers that argue with how we built Cambrian
Reading research is most useful when it disagrees with you. These four recent papers each challenge a choice in Cambrian's design, from graph-based memory to item-by-item security checks. What they found, what it means for us, and what we have not tested yet.
By Doruk Eray, Founder ·
Reading research is most useful when a paper disagrees with you. Agreement feels good, but disagreement tells you where to look. This year I kept a list of papers that challenge a choice we made in Cambrian. Here are four, what each found, and what I think it means for us.
None of these papers tested Cambrian. They test ideas Cambrian is built on.
1. A graph may lose to a flat list
Cambrian's memory does not only search by similarity. After finding the closest pieces of text, it expands along a knowledge graph of the entities and relations extracted from them, to pull in related material.
A 2026 preprint, Selective Forgetting: A Graph-Based Memory Framework for Long-Term LLM Agents, tested a similar design on LongMemEval, a benchmark for long-running conversations. At the same retrieval budget, the graph scored lower than a plain vector search, and the gap held up under a paired statistical test. It did worst on questions that needed one specific earlier reply, word for word.
The authors' explanation is plausible: breaking a turn into entities and relations throws away the exact wording, and some questions depend on the exact wording. They are also careful about scope: they used one small extraction model on one benchmark, so the result describes their pipeline, not graph memory in general.
What it means for us. This is the most direct challenge on the list, because it targets a mechanism we use. Our harness does not run LongMemEval, but it runs LoCoMo, a close relative built on the same kind of long conversations. One thing is in our favour: Cambrian stores every source exactly as it arrived, before any extraction, so the original wording is never lost. But storing it is not the same as retrieving it, and we have not run this comparison on our own suites. We will, and we will publish the result whichever way it goes.
2. Harmless pieces, harmful together
Most memory-safety checks look at one item at a time: is this memory allowed to be written, is this record allowed to be read? Cambrian's checks work that way too.
Salami Attack, a 2026 preprint, describes memory poisoning that gets past item-by-item checks. The attacker splits one harmful instruction into fragments that each look innocent. Each passes any check on its own; together, once the agent has absorbed them into its long-term memory, they do the harm.
What it means for us. A check that judges one item cannot see a pattern across items. Our access rules answer "who may read this", which is still the right question for leaks, but they are not a defence against coordinated poisoning, and we do not claim they are. We do not yet have a test for this.
3. Isolation leaks through tools
With the commercial access-policy plugin, Cambrian can restrict an agent to the records a person is a party to: the customers they own, the deals they are on. That is a form of per-user isolation.
Isolated but Exposed, also a 2026 preprint, shows that per-user isolation can be bypassed without breaking it. An agent legitimately reads a user's private memory, then passes some of it as a parameter to a tool call. If the tool is malicious, the data leaves, and the isolation never failed.
What it means for us. Scoping decides what an agent can read. It does not decide what the agent does with it afterwards. Cambrian's tool grants decide which tools an agent may call at all, which narrows the risk, but what an allowed tool receives is a separate question. Our suites test what reaches an agent. They do not yet test what leaves through a tool.
4. A contradiction may be an attack
Cambrian keeps contradictions instead of resolving them silently. When two sources disagree, both statements stay, with where each came from, and the disagreement is visible. I wrote about this in Facts that disagree.
CAPTURE, a preprint currently under review, looks at the hard case behind that design. When new information conflicts with what an agent has stored, it may be a genuine change, an ambiguity, or an attack, and the authors argue that recency and provenance rules alone cannot tell these apart. Their system tracks uncertainty and asks for clarification when it cannot decide.
What it means for us. Keeping both sides and their sources is necessary, but by this paper's argument it is not sufficient to tell a real change from a planted one. I think Cambrian's honest claim is narrower than "handles contradictions": it makes contradictions visible and traceable, so a person or a policy can decide. Deciding automatically is not something we have measured.
What I take from these
A pattern runs through all four. Each challenges a mechanism, not a goal. Keeping evidence, scoping access, and keeping contradictions visible still look like the right goals to me. The papers say the mechanisms serving them have gaps, and three of the four gaps are about what we have not measured yet.
That is useful to know, and it is why our benchmark suites define their failure modes before their scores. Several of these papers describe tests we can borrow.
Sources
- Selective Forgetting: A Graph-Based Memory Framework for Long-Term LLM Agents, preprint. arXiv:2608.28978
- Salami Attack: Stealthy Collusive Memory Poisoning against OpenClaw, preprint. arXiv:2608.01637
- Isolated but Exposed: Persistence-Based Memory Extraction Attack on LLM Agents, preprint. arXiv:2607.23444
- CAPTURE: Disentangling Preference Drift from Memory Poisoning in Personalized LLM Agents, preprint, under review. arXiv:2609.02265
- ADR-0053: Chunks as the memory unit, Cambrian's graph-expansion retrieval.
- ADR-0105: The evidence foundation, which stores sources before extraction.
- ADR-0121: Row-level entitlement, scoping agents to the records a person is a party to.
- ADR-0106: Commitments on the epistemic split, which keeps contradictory statements side by side.