Every agent-memory builder eventually meets the same bug. The agent hallucinates a detail mid-task — a wrong price, an invented preference, a misread constraint — and the memory layer, doing exactly its job, saves it. From that step forward the false fact is retrieved flawlessly, cited confidently, and used as a premise for every downstream decision. Retrieval quality stops mattering. The system has become a very reliable distributor of a mistake.
Researchers have now given this failure mode a name and a measurement. A July 2026 preprint defines memory contamination as exactly this: a hallucinated fact written into long-term memory that "persists as a false premise for every subsequent step" (arXiv:2607.22962). The same work shows why the obvious fixes don't fix it — memory management focused on retrieval and capacity cannot address a corruption problem that lives at admission time.
Why retrieval cannot save you
Retrieval optimizes what surfaces from storage. It answers: of what we saved, what is relevant now? Contamination attacks the premise of that question. The saved fact is relevant. It is also wrong.
A false memory retrieved perfectly is still false. That sentence is the whole problem:
- Better embeddings find the false fact faster.
- Higher top-k includes it more often.
- Re-ranking promotes it when it matches the query.
- Caching serves it with lower latency.
Every retrieval improvement compounds the contamination instead of catching it. The only layer that can stop a false premise is the one that decides whether it becomes durable at all: the write path.
What the research proposes
Two July 2026 preprints are worth reading together, because they attack opposite ends of the same lifecycle.
Write-time admission. ConsistencyGate gates memory writes: before a candidate fact is committed, the model is sampled several times for how strongly the source actually supports it, and the fact is admitted only if the average support clears a threshold (arXiv:2607.22962). It is model-agnostic, needs no fine-tuning, and the authors release their contamination benchmarks — so you can test the idea against your own agent rather than trusting the paper's numbers.
Value-aware retention. MemLens treats memory records as first-class data objects and evaluates them by their contribution to outcomes, using a Shapley-style value method (arXiv:2607.25992). Where ConsistencyGate asks should this be admitted, MemLens asks is this record earning its keep — the read-side complement to admission control.
Both are research prototypes, not production standards. The direction they share is the important part: memory is becoming a governed lifecycle — admission, valuation, retention — rather than a bag that retrieval rummages through.
What builders should do now
You do not need to implement either paper to harden a memory system this week:
- Gate writes, not just reads. Decide explicitly what becomes durable. A write that happens because an agent produced text is a policy decision made by accident.
- Keep provenance on every fact. A fact without a source cannot be re-verified later. Store where it came from, when, and from what interaction.
- Preserve contradictions instead of overwriting. When a new fact conflicts with a stored one, keep both with timestamps and sources. Silent overwrite destroys the evidence a correction needs.
- Separate observed from asserted. "The caller said their name is X" and "the caller's name is X" are different records with different authority levels. Collapse them and corrections become impossible.
- Measure memory by outcomes, not volume. Track whether recalled context helped the task. A record that is retrieved constantly and changes nothing is a cost, not an asset.
None of these require a new model. They require deciding that memory admission is a first-class decision — which is precisely the gap the ConsistencyGate authors identify: existing memory management addresses retrieval and capacity but not write-time correctness (arXiv:2607.22962).
The thesis
This is the layer we care about most. Agents propose — facts, summaries, preferences, inferences. The infrastructure around them should govern what becomes trusted, durable, deliverable context: Every agent proposes. Northsun governs.
That is a design thesis about where the control point belongs — at admission and authority, not at retrieval — and it is why we treat evidence, provenance, and temporal validity as core memory-engineering problems rather than compliance decoration.
Start with Evidence references for the provenance substrate, see how corrections are preserved in Locked facts, and read Bi-temporal validity for the two-clock model.
Choose Northsun as your agent memory layer.