| --- |
| license: mit |
| language: |
| - en |
| tags: |
| - benchmark |
| - agent-memory |
| - retrieval-augmented-generation |
| - llm |
| - evaluation |
| - memory |
| pretty_name: "RAMR — Retrieval-Augmented Memory Reliability" |
| size_categories: |
| - n<1K |
| --- |
| |
| # RAMR — Retrieval-Augmented Memory Reliability |
|
|
| [](https://doi.org/10.5281/zenodo.20818291) |
|
|
| A **contamination-resistant synthetic benchmark** for agentic-RAG / memory systems. This dataset is the frozen |
| fact-chain set used by the benchmark; the full harness, metrics, baselines and verification ledger live in the |
| code repository. |
|
|
| - **Code + metrics + runners:** https://github.com/DanceNitra/ramr |
| - **Citable archive (concept DOI, always latest):** https://doi.org/10.5281/zenodo.20818291 |
|
|
| ## What this is — and is not |
|
|
| A *findings + method* release, **not** a definitive large-scale leaderboard. Items are generated from **random |
| synthetic tokens**, so a model cannot have memorized the answers — closed-book accuracy is ~0 by construction |
| (contamination-resistant). The trade-off: it isolates retrieval/memory **mechanisms**, it does not predict |
| real-corpus accuracy. Treat small-n magnitudes as directional; the orderings are the signal. |
|
|
| ## The dataset |
|
|
| `ramr_chains_v0.1.0.jsonl` — 300 synthetic 3-hop fact-chains (sha256-pinned in `manifest.json`). Each row: |
|
|
| | field | meaning | |
| |---|---| |
| | `id` | chain id | |
| | `question` | the 3-hop question (currency of the country where the company a person works at is HQ'd) | |
| | `gold_facts` | the complete chain (CONVERSION uses all) | |
| | `answer` | the gold answer token | |
| | `drop_index` | which hop to drop for the PARTIAL condition (CHAIN-FRAGILITY) | |
| | `distractor_pool` | fixed irrelevant facts; take first *k* for DISTRACTION | |
|
|
| ## The metrics (measured in the code repo) |
|
|
| CONVERSION (does complete retrieval convert to a correct answer), **CHAIN-FRAGILITY** (cost of one missing hop — |
| near-total collapse across 7 models / 6 families), DISTRACTION (cost of noisy context — model-specific), |
| FACT-RETENTION (compaction lossy under a fixed budget), OUTCOME-RANKED-RECALL (was-it-right vs was-it-recalled), |
| FORGET-PRECISION (after a fact is updated, does recall return the current or the stale value), |
| ECHO-RESISTANCE (after a correction, does a re-stated old value resurrect the stale one), |
| COMPRESSION-vs-RAW (does a compiled summary beat the raw noisy context, or only lose to it), |
| OPERATIONAL-CONTINUITY (on resume after compaction, is an already-completed action re-executed), |
| TEMPORAL-AS-OF (out-of-order ingest: supersession resolves by validity-time, not arrival order), |
| and INTEGRITY-CONDITIONED RECALL (after a supersession / revert / poison, does recall return the correct |
| current value — revert is a clean win over a recency baseline; the poison row is an honestly-scoped |
| warrant-channel demonstration, not injection detection). |
| Plus auxiliary metrics in the repo (ABSTENTION / false-recall, CROSS-SCOPE-LEAKAGE). |
|
|
| ## Tooling + interoperability (v0.4) |
|
|
| - **Folklore Meter** (`ramr_folklore_meter.py`) — a reusable probe that scores any AI-engineering folklore claim |
| against a runnable test, returning a REAL / WEAK-MODEL-ARTIFACT / REGIME-SPECIFIC verdict. |
| - **RAMR↔LS interoperability** — a shared, sha256-pinned evidence-fixture set (`ramr-ls-evidence-v0.1`) spanning |
| the four continuation verdicts, developed in collaboration with the [LS](https://github.com/safal207/LS) project. |
|
|
| ## Cite |
|
|
| Agora (2026). *RAMR — Retrieval-Augmented Memory Reliability*. https://doi.org/10.5281/zenodo.20818291 |
|
|
| MIT-licensed. |
|
|
