Evaluation methodology
and results.

EDN is benchmarked against LongMemEval (Wu et al., ICLR 2025), a peer-reviewed benchmark for long-term conversational memory developed independently of EDN. Every result on this page is a measured outcome from the evaluation programme. No projections, no estimates, no modification to the benchmark protocol.

View on GitHubPreprint, Q2 2026
LongMemEval · ICLR 2025 benchmark
Evaluation snapshotMeasured
87.0%best run · Claude Haiku 4.5
Full results and methodology below.
Why memory fails

From stateless models, to reliable memory.

Two failure modes: loss and fabrication.

Language models are stateless. Across sessions, applications, and agents, prior context, corrections, and decisions are lost the moment an interaction ends. Where context is missing, the model fabricates to fill the gap, and without provenance that fabrication is written to memory and returned later as established fact.

Memory quality is a property of the system.

Whether recall is reliable, factual, and traceable is determined by the memory system, not the model alone. What is written, how each record is classified by origin, what is eligible for retrieval, and how it is assembled into context are what keep recall accurate rather than confidently wrong.

Measured, unmodified.

EDN is external, deterministic memory middleware, with no changes to model weights. We evaluated it on a public, peer-reviewed benchmark under its own protocol, introducing each architectural decision one at a time and measuring it against the configuration before it.

Results

Measured outcomes.

Best run and replication, reported together

EDN was evaluated on LongMemEval across three actor models. The primary configuration was replicated three times; we report the best run and the replication mean together, because a single good run is not evidence of a stable system.

EDN results by actor model
Actor model
Best run
Replication (n=3, mean ± σ)
Claude Haiku 4.5
87.0%
85.3% ± 0.6pp
gpt-5-mini
85.0%
single run
GPT-4o
84.2%
81.7% ± 0.50pp

Every gain is an architecture decision

Each decision was introduced one at a time and measured against the configuration before it; the model never changed. From a no-memory baseline, every layer of the system adds measured accuracy, while the provenance invariant holds throughout.

Full sub-category breakdowns, the n=3 replicated results, and ablation detail are in the paper and the public evaluation repository.

Haiku 4.5, best run only. Each bar is the running accuracy after that step; the gains sum from the 40.4% baseline to 87.0%. Across every construction the provenance invariant held, a structural guarantee separate from the accuracy shown here.

Configuration progression
Cumulative accuracy
100806040200
40.4%
Baselineno memory
+26.6pp
GovernedStructure
+16.0pp
Round-levelmemory
+4.0pp
Diversityre-ranking
87.0%
Finalfull system
Claude Haiku 4.5 Single Best Run
Benchmark LongMemEval: 500 questions over ~57M tokens of conversation history, six question categories. · Actors Claude Haiku 4.5, gpt-5-mini, GPT-4o. · Replication primary configuration n=3. · Scoring the benchmark's own LLM-judge protocol, unmodified.
Scope

What we claim, and what we don't.

Everything below is either explicitly supported by our evaluation or intentionally left outside its scope.

Supported by our evaluation
Accuracy
Measured on LongMemEval across three actor models. Best run 87.0%.
Reproducibility
Primary configuration replicated three times and reported as mean ± σ.
Structural Invariant
Zero AI_GENERATED records entered session context across all 1,500 constructions.
Retrieval Overhead
Session-package construction p95 under 50ms on every run.
Outside the scope
A leaderboard position. Published cross-system numbers mix actor models, harnesses, and reporting conventions; we measure against our own ablations instead.
Generality beyond one benchmark - a second benchmark is on the roadmap.
Large-sample statistics. n=3 is a reproducibility envelope, not a distribution.
Judge independence. Scoring uses the benchmark's single-judge protocol.
That the provenance invariant is unbreakable. It is verified as structural under evaluation conditions; adversarial stress-testing is future work.
What's next

What's next.

Active now
Typed-channel user calibration - results will be reported when that ablation is complete.
Planned
A second benchmark (LoCoMo)
Planned
An adversarial stress-test of the provenance invariant
Planned
Larger replication counts
Planned
Same-harness baselines - so comparisons stop depending on other people's reporting

Go deeper into the research.

Explore the full evaluation code, read the preprint once it's public, or reach out with methodology questions.

Evaluation repo

Explore the full evaluation code and methodology on GitHub.

View on GitHub

Preprint

The write-up is in progress and pending peer review.

Pending review

Research enquiries

Questions about the methodology or results? Reach out directly.

Contact us