diff --git a/src/benchmarks/README.md b/src/benchmarks/README.md
index 8d43cf9..4e26d1a 100644
--- a/src/benchmarks/README.md
+++ b/src/benchmarks/README.md
@@ -36,6 +36,7 @@ interface Benchmark {
 | `locomo` | GitHub snap-research/locomo | Long context memory benchmark |
 | `longmemeval` | HuggingFace xiaowu0162/longmemeval-cleaned | Long-term memory evaluation |
 | `convomem` | HuggingFace Salesforce/ConvoMem | Conversational memory benchmark |
+| `babilong` | HuggingFace RMT-team/babilong | Long-context retrieval and reasoning benchmark |
 
 ## Question Types
 
diff --git a/src/benchmarks/babilong/README.md b/src/benchmarks/babilong/README.md
new file mode 100644
index 0000000..858f885
--- /dev/null
+++ b/src/benchmarks/babilong/README.md
@@ -0,0 +1,120 @@
+# BABILong
+
+This benchmark adapter loads the BABILong evaluation dataset from local JSON files and feeds it into the standard MemoryBench pipeline.
+
+## Data layout
+
+Download the BABILong dataset from Hugging Face:
+
+https://huggingface.co/datasets/RMT-team/babilong
+
+Extract the dataset so the directory structure looks like:
+
+```text
+data/benchmarks/babilong/
+  qa1/
+    0k.json
+  qa2/
+    0k.json
+  qa3/
+    0k.json
+  ...
+  qa20/
+    0k.json
+```
+
+The adapter currently loads the `0k.json` evaluation split from every `qa1`–`qa20` task.
+
+Each task contains 100 evaluation samples.
+
+## Dataset format
+
+Each sample follows the native BABILong format:
+
+```json
+{
+  "input": "John travelled to the hallway. Mary journeyed to the bathroom...",
+  "question": "Where is Mary?",
+  "target": "bathroom"
+}
+```
+
+The adapter maps each sample to:
+
+- one `UnifiedSession` containing the document (`input`)
+- one benchmark question (`question`)
+- one ground-truth answer (`target`)
+
+## List questions
+
+```bash
+bun run src/index.ts list-questions -b babilong -l 10
+```
+
+Filter by task:
+
+```bash
+bun run src/index.ts list-questions -b babilong -t qa7 -l 10
+```
+
+## Ingest into a provider
+
+Example using Supermemory:
+
+PowerShell:
+
+```powershell
+$env:SUPERMEMORY_API_KEY = "sm_xxx"
+bun run src/index.ts ingest -p supermemory -b babilong -r babilong-smoke --force
+```
+
+Bash:
+
+```bash
+SUPERMEMORY_API_KEY=sm_xxx \
+bun run src/index.ts ingest \
+  -p supermemory \
+  -b babilong \
+  -r babilong-smoke \
+  --force
+```
+
+## Search after ingest
+
+```bash
+bun run src/index.ts search -r babilong-smoke
+```
+
+Search results are written to:
+
+```text
+data/runs/babilong-smoke/results/
+```
+
+## What gets ingested
+
+For every BABILong sample, the adapter creates:
+
+- one session containing the complete document (`input`)
+- one benchmark question (`question`)
+- one ground-truth answer (`target`)
+
+This preserves the original long-context retrieval task while using the standard MemoryBench evaluation pipeline.
+
+## Benchmark
+
+BABILong evaluates long-context retrieval and reasoning by embedding a small set of supporting facts inside large amounts of irrelevant text.
+
+The benchmark contains 20 reasoning tasks (`qa1`–`qa20`) covering skills such as:
+
+- single supporting fact
+- multi-hop reasoning
+- counting
+- negation
+- coreference
+- deduction
+- induction
+- positional reasoning
+- path finding
+
+The adapter keeps the original task names as MemoryBench question types, allowing individual reasoning capabilities to be benchmarked independently.
\ No newline at end of file
diff --git a/src/benchmarks/babilong/index.ts b/src/benchmarks/babilong/index.ts
new file mode 100644
index 0000000..6017fcc
--- /dev/null
+++ b/src/benchmarks/babilong/index.ts
@@ -0,0 +1,153 @@
+import { existsSync, readFileSync } from "fs"
+import { join } from "path"
+import type { Benchmark, BenchmarkConfig, QuestionFilter } from "../../types/benchmark"
+import type {
+  UnifiedQuestion,
+  UnifiedSession,
+  UnifiedMessage,
+  QuestionTypeRegistry,
+} from "../../types/unified"
+import type { BABILongItem } from "./types"
+import { logger } from "../../utils/logger"
+
+const DEFAULT_DATA_PATH = "./data/benchmarks/babilong"
+const DEFAULT_CONTEXT_LENGTH = "0k"
+
+// qa subfolders (all use 0k format)
+const QA_FOLDERS = Array.from({ length: 20 }, (_, i) => `qa${i + 1}`)
+
+/**
+ * BABILong question types - native task types from the dataset.
+ */
+export const BABILONG_QUESTION_TYPES: QuestionTypeRegistry = {
+  qa1: { id: "qa1", alias: "single", description: "Single supporting fact" },
+  qa2: { id: "qa2", alias: "two", description: "Two supporting facts" },
+  qa3: { id: "qa3", alias: "three", description: "Three supporting facts" },
+  qa4: { id: "qa4", alias: "two-arg", description: "Two argument relations" },
+  qa5: { id: "qa5", alias: "three-arg", description: "Three argument relations" },
+  qa6: { id: "qa6", alias: "yes-no", description: "Yes/no questions" },
+  qa7: { id: "qa7", alias: "counting", description: "Counting" },
+  qa8: { id: "qa8", alias: "lists-sets", description: "Lists/sets" },
+  qa9: { id: "qa9", alias: "simple-neg", description: "Simple negation" },
+  qa10: { id: "qa10", alias: "indef-know", description: "Indefinite knowledge" },
+  qa11: { id: "qa11", alias: "basic-coref", description: "Basic coreference" },
+  qa12: { id: "qa12", alias: "conjunction", description: "Conjunction" },
+  qa13: { id: "qa13", alias: "compound-coref", description: "Compound coreference" },
+  qa14: { id: "qa14", alias: "time-reasoning", description: "Time reasoning" },
+  qa15: { id: "qa15", alias: "basic-deduction", description: "Basic deduction" },
+  qa16: { id: "qa16", alias: "basic-induction", description: "Basic induction" },
+  qa17: { id: "qa17", alias: "positional-reasoning", description: "Positional reasoning" },
+  qa18: { id: "qa18", alias: "size-reasoning", description: "Size reasoning" },
+  qa19: { id: "qa19", alias: "path-finding", description: "Path finding" },
+  qa20: { id: "qa20", alias: "qa20", description: "Task 20" },
+}
+
+export class BABILongBenchmark implements Benchmark {
+  name = "babilong"
+
+  private questions: UnifiedQuestion[] = []
+  private sessionsMap: Map<string, UnifiedSession[]> = new Map()
+  private dataPath: string = ""
+
+  async load(config?: BenchmarkConfig): Promise<void> {
+    this.dataPath = config?.dataPath || DEFAULT_DATA_PATH
+    const fullPath = join(process.cwd(), this.dataPath)
+
+    if (!existsSync(fullPath)) {
+      throw new Error(
+        `BABILong dataset not found at ${fullPath}. Download the dataset from https://huggingface.co/datasets/RMT-team/babilong and place the qa1-qa20 folders under ${this.dataPath}.`
+      )
+    }
+
+    this.loadQuestions(fullPath)
+  }
+
+  private loadQuestions(fullPath: string): void {
+    for (const qaFolder of QA_FOLDERS) {
+      const filePath = join(fullPath, qaFolder, `${DEFAULT_CONTEXT_LENGTH}.json`)
+
+      if (!existsSync(filePath)) {
+        logger.warn(`Missing file for ${qaFolder}: ${filePath}`)
+        continue
+      }
+
+      try {
+        const content = readFileSync(filePath, "utf8")
+        const items: BABILongItem[] = JSON.parse(content)
+
+        for (let i = 0; i < items.length; i++) {
+          this.processItem(items[i], qaFolder, i)
+        }
+      } catch (e) {
+        logger.error(`Failed to load BABILong data for ${qaFolder}: ${e}`)
+      }
+    }
+
+    logger.info(`Loaded ${this.questions.length} questions from BABILong`)
+  }
+
+  private processItem(item: BABILongItem, qaFolder: string, index: number): void {
+    const questionId = `babilong-${qaFolder}-${String(index).padStart(3, "0")}`
+
+    const session = this.extractSession(item, questionId)
+
+    this.questions.push({
+      questionId,
+      question: item.question,
+      questionType: qaFolder,
+      groundTruth: item.target,
+      haystackSessionIds: [session.sessionId],
+      metadata: {
+        task: qaFolder,
+        contextLength: DEFAULT_CONTEXT_LENGTH,
+      },
+    })
+
+    this.sessionsMap.set(questionId, [session])
+  }
+
+  private extractSession(item: BABILongItem, questionId: string): UnifiedSession {
+    const message: UnifiedMessage = {
+      role: "user",
+      content: item.input,
+    }
+
+    return {
+      sessionId: `${questionId}-session-0`,
+      messages: [message],
+    }
+  }
+
+  getQuestions(filter?: QuestionFilter): UnifiedQuestion[] {
+    let result = [...this.questions]
+
+    if (filter?.questionTypes?.length) {
+      result = result.filter((q) => filter.questionTypes!.includes(q.questionType))
+    }
+
+    if (filter?.offset) {
+      result = result.slice(filter.offset)
+    }
+
+    if (filter?.limit) {
+      result = result.slice(0, filter.limit)
+    }
+
+    return result
+  }
+
+  getHaystackSessions(questionId: string): UnifiedSession[] {
+    return this.sessionsMap.get(questionId) || []
+  }
+
+  getGroundTruth(questionId: string): string {
+    const question = this.questions.find((q) => q.questionId === questionId)
+    return question?.groundTruth || ""
+  }
+
+  getQuestionTypes(): QuestionTypeRegistry {
+    return BABILONG_QUESTION_TYPES
+  }
+}
+
+export default BABILongBenchmark
\ No newline at end of file
diff --git a/src/benchmarks/babilong/types.ts b/src/benchmarks/babilong/types.ts
new file mode 100644
index 0000000..d0bb451
--- /dev/null
+++ b/src/benchmarks/babilong/types.ts
@@ -0,0 +1,5 @@
+export interface BABILongItem {
+  input: string
+  question: string
+  target: string
+}
\ No newline at end of file
diff --git a/src/benchmarks/index.ts b/src/benchmarks/index.ts
index b790e87..5848744 100644
--- a/src/benchmarks/index.ts
+++ b/src/benchmarks/index.ts
@@ -2,11 +2,13 @@ import type { Benchmark, BenchmarkName } from "../types/benchmark"
 import { LoCoMoBenchmark } from "./locomo"
 import { LongMemEvalBenchmark } from "./longmemeval"
 import { ConvoMemBenchmark } from "./convomem"
+import { BABILongBenchmark } from "./babilong"
 
 const benchmarks: Record<BenchmarkName, new () => Benchmark> = {
   locomo: LoCoMoBenchmark,
   longmemeval: LongMemEvalBenchmark,
   convomem: ConvoMemBenchmark,
+  babilong: BABILongBenchmark,
 }
 
 export function createBenchmark(name: BenchmarkName): Benchmark {
@@ -21,4 +23,4 @@ export function getAvailableBenchmarks(): BenchmarkName[] {
   return Object.keys(benchmarks) as BenchmarkName[]
 }
 
-export { LoCoMoBenchmark, LongMemEvalBenchmark, ConvoMemBenchmark }
+export { LoCoMoBenchmark, LongMemEvalBenchmark, ConvoMemBenchmark, BABILongBenchmark }
diff --git a/src/cli/index.ts b/src/cli/index.ts
index b3c29d6..89f33d5 100644
--- a/src/cli/index.ts
+++ b/src/cli/index.ts
@@ -150,11 +150,16 @@ Available benchmark datasets for evaluation:
   convomem       ConvoMem - Conversational memory benchmark
                  Tests: user facts, assistant facts, preferences, implicit connections
                  Source: HuggingFace Salesforce/ConvoMem (downloaded on first use)
+  
+  babilong      BABILong - Long-context retrieval benchmark
+                Tests: retrieval, multi-hop reasoning, deduction, counting, coreference
+                Source: HuggingFace RMT-team/babilong (local data required)
 
 Usage:
   -b locomo        Run LoCoMo benchmark
   -b longmemeval   Run LongMemEval benchmark
   -b convomem      Run ConvoMem benchmark
+  -b babilong      Run BABILong benchmark
 `)
 }
 
diff --git a/src/types/benchmark.ts b/src/types/benchmark.ts
index 07961e4..fe53c68 100644
--- a/src/types/benchmark.ts
+++ b/src/types/benchmark.ts
@@ -20,4 +20,4 @@ export interface Benchmark {
   getQuestionTypes(): QuestionTypeRegistry
 }
 
-export type BenchmarkName = "locomo" | "longmemeval" | "convomem"
+export type BenchmarkName = "locomo" | "longmemeval" | "convomem" | "babilong"