Supersession
Cascade
A changed fact gets a supersedes edge and the old value is marked superseded. Dependency rules walk the graph, so facts derived from the changed one update — or are flagged — in the same write.
The three unsolved problems of agent memory · MEME (KAIST 2026)
The field stores facts. GRIFFai revises beliefs.
The independent MEME benchmark names three problems the entire field has not solved — Cascade, Absence, Deletion — and finds the field under 6% where it matters most. GRIFFai broke the curve on the full 100-episode benchmark, with MEME’s own judges. Not an increment. A different architecture.
18×
Cascade
.561 vs the published field average of ~.03
60×
Absence
.600 vs the published field average of ~.01
3.3×
Deletion
.890 vs the best published system at ~.27
100
Episodes
Six task types, scored by MEME’s own judges
Complete scorecard · 100 episodes · MEME judges
Same answer model. Same judges. Different data model.
The decisive control: GRIFFai ran the identical benchmark twice — once as a plain retrieval store, the way most memory products operate, and once with its integrity primitives engaged. The plain store fails like the field. The governed configuration does not. The entire gap is the architecture.
| Configuration | Cascade | Absence | Deletion | Tracking | Exact recall | Aggregation |
|---|---|---|---|---|---|---|
| Published field average | ~.03 | ~.01 | — | — | — | — |
| Best published system | ~.06 | ~.05 | ~.27 | — | — | — |
| GRIFFai as plain store (control) | .085 | .008 | .280 | .760 | .990 | .440 |
| GRIFFai governed — max integrity | .561 | .600 | .890 | .840 | .950 | .150 |
| GRIFFai governed — balanced default | .500 | .592 | .820 | .820 | .970 | .290 |
GRIFFai’s reproduction of the public MEME benchmark: 100 episodes, six task types, MEME’s published judge prompts and models — not an official leaderboard submission. The harness was calibrated by reproducing the paper’s baseline row before any GRIFFai configuration was scored. Published results are single-seed with a single answer-LLM/judge pair (gpt-4.1-mini / gpt-4o), matching the paper’s main configuration; a multi-seed replication with cost-per-episode figures is queued as follow-up work and will be published on this page. Field figures from the MEME paper, arXiv:2605.12477.
The mechanism
Belief revision happens when the fact changes — not when the question arrives.
Most systems store everything and hope the answering model sorts out contradictions at read time. GRIFFai’s Brain resolves them at write time, so retrieval serves an already-consistent state. MEME’s three unsolved tasks map one-to-one onto three GRIFFai integrity primitives.
Staleness
Absence
When an upstream fact changes and no replacement was stated, dependents become uncertain. The system reports what it no longer knows instead of confidently serving a stale value.
Tombstone
Deletion
“Forget that” marks the record tombstoned: the value is withheld from every recall while the record’s existence remains auditable. Deletion is honored, not just filtered.
Benchmark release · August 2026
Read the full report.
Method, complete scorecard, the four-configuration ablation study — including the configurations that did not work — and a verification appendix with hashes and negative-control probes. Nine pages.
Read the full report (PDF) →Why it matters in production
A vendor changes. A policy updates. A price moves. Every fact derived from that change has to move with it — and an agent on store-everything memory keeps acting on the old chain. Each MEME task is a production failure mode; these scores measure how often it is prevented.
The novel part
The field keeps bolting smarter retrieval onto store-everything memory. GRIFFai moved belief revision into the write path — supersession edges, tombstones, staleness states, dependency cascade. Memory integrity is a data-model property: engineered at write time, or it does not exist.
Source
MEME: an independent benchmark for episodic memory with belief revision, KAIST 2026. Read the paper on arXiv →
In the paper’s own words
Don’t take our word for the problem. Take theirs.
MEME was authored by researchers at KAIST AI, the Tübingen AI Center, and NAVER AI Lab. Read what the paper itself concludes about the field — then compare it with the scorecard above.
The field’s verdict
“All systems collapse on dependency reasoning… every practical-cost configuration fails Cascade and Absence.”
Six memory systems, three architectural paradigms, 100 controlled episodes — and prompt optimization, deeper retrieval, reduced noise, and stronger LLMs all fail to close the gap.
The cost wall
“This configuration costs ~70× the baseline and is not deployable today.”
The paper’s only partial fix — a file agent driven by a frontier LLM — reaches Cascade .32 and Absence .59 at ~70× baseline cost. GRIFFai scores .561 and .600 with the same mini-model class the paper uses as its 1× cost baseline.
The architecture they call for
“The path forward is memory architectures that natively propagate updates through dependent facts at maintenance… We leave the architecture open.”
That is a published description of GRIFFai’s write path — supersession, staleness, tombstones, and dependency cascade, shipping in GRIFFai’s data model before the paper named the need.
The paper
MEME: Multi-entity & Evolving Memory Evaluation
Jung, Rubinstein, Uselis, Yun & Oh — KAIST AI · Tübingen AI Center · NAVER AI Lab, 2026. Hosted here for convenience; canonical version on arXiv. Quotes above are from the abstract and conclusion; the 70× comparison uses the paper’s Table 4 internal-LLM ablation (20-episode subset).
Evidence pack · verify it yourself
Don’t trust the summary. Re-score the answers.
The complete per-question record of every published run — 7,048 records: question, gold answer, our system’s answer, the judge’s verdict and rationale — is downloadable, with per-record SHA-256 hashes chained to the report’s verification appendix. Every number on this page is a deterministic function of those records.
Recompute
Every aggregate
All six runs — both controls, both governed configurations, and the two ablations that did not work — as per-question JSONL. Sum the pass fields and compare against the scorecard above.
Re-judge
With your own key
The included rejudge.py re-runs MEME’s verbatim gpt-4o judge prompts over our published answers under your own API key and reports agreement — no GRIFFai access required.
Reproduce
Against any system
The pack ships the full scoring harness — verbatim judge prompts, chunking, protocol — with a pluggable memory interface (Apache-2.0), so any memory system can be scored under identical conditions.
Independent evaluation program
Run it yourself, against us.
Named evaluators — researchers, customers under evaluation, technical press — run the public MEME benchmark against a sealed GRIFFai endpoint using the pack’s included adapter: your dataset copy, your seed, your judge, your key. You observe behavior, not internals, and you publish your numbers without our approval.
Next step
Put governed memory under your agents.
The integrity primitives measured here are the same supersession, staleness, and tombstone machinery in GRIFFai’s data model. Their MEME-relevant invariant — a superseded value never reaches the positive-fact lane — is enforced and probed at the serving boundary of the production front door; the boundary-probe methodology and results are in the report’s verification appendix. See what that looks like against your own workflows.