agentic-recall-v1

112,740 multi-turn agentic trajectories augmented with a "recall" turn — late in the conversation the user asks the assistant to recall a fact or action established much earlier, and the assistant answers from memory, with no new tool call. Built to strengthen long-context recall (the failure mode that shows up after context compaction).

Every row is an existing trajectory with one recall exchange added (user question + assistant reply), appended at the end or inserted mid-conversation. The recalled fact is always rule-verified, so the dataset never teaches an incorrect recall. The original rows are preserved; these are separate -recall rows and are meant to be mixed with the base corpora during training, not used alone.

Subsets

config	rows	base	language	how the gold is fixed
appworld	5,679	agentic-selfgen-ko-v1	ko	a rule computes the answer (count / list / sum / last of the episode's successful actions)
nemotron-ko	43,133	agentic-think-v2-react-hybrid (nemotron, translated)	ko	27B extracts a verbatim fact; a rule verifies it appears in an earlier turn (grounding)
nemotron-en	63,928	agentic-think-v2-react-hybrid (nemotron)	en	same as above, in English
total	112,740

Two mechanisms, same guarantee — the recalled value is provably correct:

• appworld has a fixed 4-app schema (phone / notes / tasks / payments), so a rule reads the successful tool calls and computes the answer: "how many notes did I create", "what did I send", "total transferred", "what was the last alarm". Answers are rule-generated; teacher 27B paraphrases the question for variety.
• nemotron has ~840 arbitrary tool schemas, so a rule cannot compute rich facts. Instead the 27B picks a short concrete fact that appeared verbatim earlier (an id, code, email, amount, date, name, result), and a rule drops the row unless that fact really occurs in an earlier turn. This blocks hallucinated recall (2.2% of en candidates were dropped this way).

What a recall turn looks like

(… 20 turns of agentic tool use …)
user:      방금 이메일로 받으신 인증 코드가 무엇이었나요?
assistant: (reasoning) 사용자가 앞선 대화에 나온 정보를 다시 확인한다. 앞서 587291이 나왔으니 그대로 알려 준다.
           (content)   인증 코드는 587291입니다.

user:      What was the output of the factorial function for the input 5?
assistant: (reasoning) The user is asking me to recall information from earlier …
           (content)   The function returned 120 for the input 5.

The recall question sits a median of 6–7 turns (up to ~37) after the fact first appeared. The assistant answer is plain text (no markdown) and keeps the reasoning_content visible, so training can mask or keep the reasoning on the recall turn as it chooses.

Schema

field	type	description
id	str	base id + -recall
source	str	base source (kotraj_appworld_*, nemotron_tc, nemotron_ia)
tools	str (JSON)	the base episode's tool catalog
messages	str (JSON)	base conversation with the recall user/assistant turns added; the recall assistant turn has content + reasoning_content and no tool_calls
think_source	str	recall_rule / recall_27b (appworld), recall_nm_ko / recall_nm_en (nemotron)
meta	str (JSON)	recall: the gold fact/answer, source turn, recall distance, placement (end/mid)

from datasets import load_dataset
ds = load_dataset("HBKenerzai/agentic-recall-v1", "nemotron-ko", split="train")

Quality

Independently re-verified on 3,000-row samples per subset with 0 issues: the recalled fact genuinely appears in an earlier turn, it is not leaked in the question, it is present in the answer, and the answer contains no markdown.

Access

This dataset is gated: access is granted on request after manual review.

Provenance

Built by ENERZAi with the agentic-datagen data_gen/kotraj_recall.py, gen_recall.py, gen_nemotron_recall.py. Teacher for paraphrase/extraction: Qwen3.8-27B-FP8. Bases: AppWorld-derived self-generated trajectories and the nemotron portion of agentic-think-v2-react-hybrid. Related: agentic-think-v2-react-hybrid-ko, agentic-selfgen-ko-v1.