{"paper":{"title":"PREPING: Building Agent Memory without Tasks","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Agents can construct competitive procedural memory for new environments using only self-generated synthetic tasks before any real experience.","cross_cats":["cs.CL"],"primary_cat":"cs.AI","authors_text":"Jinheon Baek, Minki Kang, Sangwoo Park, Sung Ju Hwang, Yumin Choi","submitted_at":"2026-05-11T04:34:43Z","abstract_excerpt":"Agent memory is typically constructed either offline from curated demonstrations or online from post-deployment interactions. However, regardless of how it is built, an agent faces a cold-start gap when first introduced to a new environment without any task-specific experience available. In this paper, we study pre-task memory construction: whether an agent can build procedural memory before observing any target-environment tasks, using only self-generated synthetic practice. Yet, synthetic interaction alone is insufficient, as without controlling what to practice and what to store, synthetic "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Preping substantially improves over a no-memory baseline and achieves performance competitive with strong playbook-based methods built from offline or online experience, with deployment cost $2.99× lower on AppWorld and $2.23× lower on BFCL v3 than online memory construction.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That synthetic tasks generated and filtered by the proposer-validator loop will transfer to real target-environment tasks without any direct experience of those tasks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Preping builds agent memory via proposer-guided synthetic practice and selective validation, matching offline/online methods at 2-3x lower deployment cost.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Agents can construct competitive procedural memory for new environments using only self-generated synthetic tasks before any real experience.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"6d51bf3eb1a40f87addfa0a6437e69e7717fa2c8358465e2b5b4b19fce4c94d6"},"source":{"id":"2605.13880","kind":"arxiv","version":1},"verdict":{"id":"11a37770-dc45-4a86-9c31-d818ca02939b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T06:10:55.020382Z","strongest_claim":"Preping substantially improves over a no-memory baseline and achieves performance competitive with strong playbook-based methods built from offline or online experience, with deployment cost $2.99× lower on AppWorld and $2.23× lower on BFCL v3 than online memory construction.","one_line_summary":"Preping builds agent memory via proposer-guided synthetic practice and selective validation, matching offline/online methods at 2-3x lower deployment cost.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That synthetic tasks generated and filtered by the proposer-validator loop will transfer to real target-environment tasks without any direct experience of those tasks.","pith_extraction_headline":"Agents can construct competitive procedural memory for new environments using only self-generated synthetic tasks before any real experience."},"references":{"count":53,"sample":[{"doi":"10.48550/arxiv.2602.21320","year":2026,"title":"Tool-r0: Self-evolving llm agents for tool-learning from zero data","work_id":"682365bd-b781-4d4c-a78a-a0351adbdc31","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Introducing the model context protocol, November 2024","work_id":"f0cdfef8-384f-4258-b107-3db84142d5ca","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"DeepSeek-V3.2: Pushing the frontier of open large language models, 2025","work_id":"812ba0f2-05ea-49f6-a758-c2ae6d0fc49f","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.48550/arxiv.2510.04851","year":2025,"title":"LEGOMem: Modular Procedural Memory for Multi-agent LLM Systems for Workflow Automation","work_id":"81b5a1e3-7f36-4854-a305-10e386c1b07f","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.48550/arxiv.2508.05004","year":2025,"title":"R-Zero: Self-Evolving Reasoning LLM from Zero Data","work_id":"45b39aae-1151-4087-9b48-dd16313e6306","ref_index":5,"cited_arxiv_id":"2508.05004","is_internal_anchor":true}],"resolved_work":53,"snapshot_sha256":"320ef9a362b9f83b5cba69ff837dea30c9ab8274d36be88aa5764a1597a07c53","internal_anchors":7},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}