Storing the Evidence Before the Belief

On a memory architecture that separates what an agent knows from what it was told, and what that separation exposes about the systems that do not make it.

On May 29, 2026, Resham Joshi published Eywa: Provenance-Grounded Long-Term Memory for AI Agents on arXiv, describing a memory architecture built around a specific design principle: evidence before belief. The paper's opening diagnosis is aimed squarely at how most current agent memory systems are built, and it is unsparing. Existing memory systems often collapse source evidence, extracted facts, retrieved context, and answer policy into one opaque prompt path, the paper states, making failures difficult to diagnose: a wrong answer may come from missing evidence, unsupported extraction, stale state, retrieval loss, or answer-model behavior.

That diagnosis names something the broader field has mostly treated as an acceptable tradeoff: when an agent gives a wrong answer after consulting its memory, there is typically no way to determine which of five distinct failure points produced the error, because the architecture never separated them in the first place. The evidence, the extracted fact, the retrieved context, and the policy that turned all of that into an answer are fused into a single opaque path. Diagnosing a failure means guessing among five possible causes with no artifact that isolates which one actually occurred.

The Architecture

Eywa's structural response is to keep these components separate by construction. The system stores immutable source evidence before deriving canonical facts, validates extracted memories against typed signals and source support, and retrieves bounded memory context through a deterministic multi-route read path with zero LLM calls inside retrieval. Each clause in that sentence closes off a specific failure mode the opening diagnosis named. Immutable evidence means the original source cannot be silently altered by later processing. Validating extracted memories against source support means a fact cannot enter the system's belief store without a documented chain back to what it was derived from. A deterministic retrieval path with zero LLM calls inside retrieval means the act of pulling memory into context is not itself an additional point where probabilistic behavior can introduce untracked variance.

The paper adds a further separation: retrieved context is returned separately from answer instructions, allowing the same memory substrate to be evaluated across frontier, budget, and local answer models. This means the memory layer's correctness can be assessed independently of which model is doing the reasoning on top of it, a separation that most fused architectures do not allow, because in a fused system, memory quality and answer quality are entangled in a single opaque score.

What the Numbers Show

The paper reports Eywa reaching 90.19% judge accuracy on the LoCoMo C1-C4 split using Claude Sonnet 4.6 in write and question-answering roles, under a frozen, artifact-recorded retrieval configuration. On a separate benchmark, LongMemEval-S, it reports 88.2% retrieval-sufficiency accuracy. On BEAM, described as a 700-question technical-memory stress benchmark, the paper reports 81.45% mean nugget score and 85.29% pass rate at a score threshold of 0.5. The paper states that full per-question artifacts, including questions, gold answers, model answers, retrieved context, and labels, are published for external review, a choice consistent with the paper's own emphasis on evidence that can be checked rather than results that must be taken on faith.

Provenance as the Missing Layer in the Arc

This paper closes a gap that has been implicit across the rest of this arc without being named directly. The UK AI Safety Institute's oversight degradation report identified internal activations, chain-of-thought, external actions, and inter-agent communication as the four surfaces oversight depends on, but did not address memory as a fifth surface with its own legibility problem. The Frontier Model Forum's issue brief separately warned that malicious inputs "can be stored in memory, propagate across agents, and reappear in later decision cycles," identifying memory as an attack vector without proposing an architecture that would make the propagation traceable after the fact. Eywa is, in effect, a proposed answer to the traceability half of that problem: if a wrong belief propagated from memory into a later decision, an architecture built around evidence-before-belief is designed to let an investigator trace that belief back to the specific piece of evidence, or lack of evidence, that produced it.

What a Reference Architecture Does Not Establish

A single research paper demonstrating strong benchmark performance for a provenance-grounded memory architecture is not the same as that architecture becoming the substrate any deployed agent system actually uses. The benchmarks Eywa reports, LoCoMo, LongMemEval-S, and BEAM, are research evaluation environments. None of them establishes that a production agent operating inside a live enterprise workflow, with its own memory built on a fused, opaque architecture of the kind the paper critiques, has any path to retrofitting provenance tracking after deployment. The paper demonstrates that evidence-before-belief is achievable and can perform well. It does not demonstrate that the field's existing deployed memory systems are moving toward that design.

Open Questions

• How many currently deployed agent memory systems could be audited for the five failure points, missing evidence, unsupported extraction, stale state, retrieval loss, answer-model behavior, that Eywa's architecture is designed to distinguish?
• What would it take to retrofit provenance tracking into an agent system whose memory architecture already fuses evidence, extraction, retrieval, and policy into a single opaque path?
• At what point does an enterprise deploying agents with persistent memory become obligated to demonstrate that a wrong output can be traced to a specific evidentiary failure, rather than reported as an unexplained error?
• How does claim-to-source lineage survive when the same memory substrate is queried by multiple different answer models with different reasoning behaviors, as the paper's cross-model evaluation implies is possible?
• Who verifies that a provenance architecture's own validation step, checking extracted memories against typed signals and source support, is itself functioning correctly rather than passing unsupported claims through?
• Does publishing full per-question artifacts for external review, as this paper does, become an expectation for provenance-memory claims generally, or remain the exception?

The loop closed around an oversight function that was never instrumented.