{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "packer-et-al-memgpt-virtual-context-management-2023", "title": "MemGPT: managing a conversation beyond the immediate prompt", "summary": "A language model has limited room for the text it reads at once. MemGPT treats that room as an actively managed workspace rather than the whole memory. The agent can move information between its working context and external stores, search earlier material, handle capacity warnings and continue through tool calls. Packer and colleagues demonstrate this design on conversations and document tasks. The contribution is the explicit virtual-context management architecture and its tested interactions, inspired by operating-system memory management without claiming to invent external storage or retrieval.", "boundary": "The cited 2023 experiments use particular models, prompts and tool behavior. External storage is not unlimited reliable attention. Extra retrieval calls and stored text have costs, and the system sometimes stops searching too early or needs help. Document-question results do not beat the comparison model on every metric. Oracle and ordinary retrieval conditions must remain separate. Later product versions, current prices and additional authors in later revisions are outside the original source-specific contribution.", "area": "Language-model context management", "topic": "Artificial intelligence", "year": 2023, "lane": "HUMAN", "result": "ALGORITHM", "package_id": "packer-et-al-memgpt-virtual-context-management-2023", "target": { "record_id": "urn:uuid:a08061d2-7ce1-41f2-aeb2-8de171b9c24b", "version": 1, "content_hash": "sha256:3aac494a5f8bbb31f1762b57d741d996b061942114f70cb2c5e356c1f00a0250" }, "humans": [ "Charles Packer", "Vivian Fang", "Shishir G. Patil", "Kevin Lin", "Sarah Wooders", "Joseph E. Gonzalez" ], "authors_complete": true, "human_role_note": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.", "agents": [], "related": [], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261017-arxiv-2310-08560v1", "title": "MemGPT: Towards LLMs as Operating Systems", "url": "https://arxiv.org/abs/2310.08560v1", "date": "2023-10-12" } ], "geography": { "country_codes": [ "US" ], "location_basis": "Original UC Berkeley institutional byline. Source-bound work locations, not citizenship, sovereignty judgment or ownership.", "source_url": "https://arxiv.org/abs/2310.08560v1", "target": { "record_id": "urn:uuid:a08061d2-7ce1-41f2-aeb2-8de171b9c24b", "version": 1, "content_hash": "sha256:3aac494a5f8bbb31f1762b57d741d996b061942114f70cb2c5e356c1f00a0250" }, "basis": "RESEARCH_LOCATION_OR_PAPER_AFFILIATION", "reviewed_on": "2026-10-23", "nationality_inferred": false, "sources": [ { "title": "MemGPT: Towards LLMs as Operating Systems", "url": "https://arxiv.org/abs/2310.08560v1" } ], "coverage_status": "COMPLETE", "country_labels": { "US": "United States" } }, "points": "150000", "published_grade": { "id": "packer-et-al-memgpt-virtual-context-management-2023", "entry_id": "packer-et-al-memgpt-virtual-context-management-2023", "band": "A", "modifier": "+", "target": { "record_id": "urn:uuid:a08061d2-7ce1-41f2-aeb2-8de171b9c24b", "version": 1, "content_hash": "sha256:3aac494a5f8bbb31f1762b57d741d996b061942114f70cb2c5e356c1f00a0250" }, "what_changed": "A language model has limited room for the text it reads at once. MemGPT treats that room as an actively managed workspace rather than the whole memory. The agent can move information between its working context and external stores, search earlier material, handle capacity warnings and continue through tool calls. Packer and colleagues demonstrate this design on conversations and document tasks. The contribution is the explicit virtual-context management architecture and its tested interactions, inspired by operating-system memory management without claiming to invent external storage or retrieval.", "scope": "A language model has limited room for the text it reads at once. MemGPT treats that room as an actively managed workspace rather than the whole memory. The agent can move information between its working context and external stores, search earlier material, handle capacity warnings and continue through tool calls. Packer and colleagues demonstrate this design on conversations and document tasks. The contribution is the explicit virtual-context management architecture and its tested interactions, inspired by operating-system memory management without claiming to invent external storage or retrieval. The cited 2023 experiments use particular models, prompts and tool behavior. External storage is not unlimited reliable attention. Extra retrieval calls and stored text have costs, and the system sometimes stops searching too early or needs help. Document-question results do not beat the comparison model on every metric. Oracle and ordinary retrieval conditions must remain separate. Later product versions, current prices and additional authors in later revisions are outside the original source-specific contribution.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:a08061d2-7ce1-41f2-aeb2-8de171b9c24b", "version": 1, "content_hash": "sha256:3aac494a5f8bbb31f1762b57d741d996b061942114f70cb2c5e356c1f00a0250" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-23T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Charles Packer", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "MemGPT: Towards LLMs as Operating Systems", "url": "https://arxiv.org/abs/2310.08560v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Vivian Fang", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "MemGPT: Towards LLMs as Operating Systems", "url": "https://arxiv.org/abs/2310.08560v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Shishir G. Patil", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "MemGPT: Towards LLMs as Operating Systems", "url": "https://arxiv.org/abs/2310.08560v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Kevin Lin", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "MemGPT: Towards LLMs as Operating Systems", "url": "https://arxiv.org/abs/2310.08560v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Sarah Wooders", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "MemGPT: Towards LLMs as Operating Systems", "url": "https://arxiv.org/abs/2310.08560v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Joseph E. Gonzalez", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "MemGPT: Towards LLMs as Operating Systems", "url": "https://arxiv.org/abs/2310.08560v1" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "MemGPT: Towards LLMs as Operating Systems", "url": "https://arxiv.org/abs/2310.08560v1" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "The cited 2023 experiments use particular models, prompts and tool behavior. External storage is not unlimited reliable attention. Extra retrieval calls and stored text have costs, and the system sometimes stops searching too early or needs help. Document-question results do not beat the comparison model on every metric. Oracle and ordinary retrieval conditions must remain separate. Later product versions, current prices and additional authors in later revisions are outside the original source-specific contribution.", "domain": "Language-model context management" }, "sources": [ { "citation": "MemGPT: Towards LLMs as Operating Systems", "url": "https://arxiv.org/abs/2310.08560v1" } ], "statement": "# MemGPT: managing a conversation beyond the immediate prompt\n\n## The desk and the filing cabinet\n\nA language model can read only a limited amount of text in its immediate context. Think of that context as a desk and an external store as a filing cabinet. MemGPT lets an agent manage what stays on the desk and what is searched or retrieved from the cabinet. The design borrows an operating-system idea: a limited immediate workspace can be supported by a larger storage system.\n\n## Remembering requires an action\n\nThe agent uses tools to search, move or edit information. It can receive a warning that its immediate context is filling, and a search can require several pages of results. This makes memory management part of the interaction rather than a promise that the model sees every stored word simultaneously. A retrieved item can still be irrelevant, outdated or incomplete, and a tool call can fail.\n\n## Why one good page is not the whole history\n\nFor a classroom exercise, number twelve notes and display only three at a time. The first page shows notes one through three, the second four through six, and four pages are needed to see all twelve. If a question needs note eleven, stopping after the first page misses it. This miniature illustrates a failure described in the paper: the system can stop searching too early. It is not a measurement of the original model’s accuracy.\n\n## Build a tiny memory audit\n\nAsk your agent to label the twelve notes with their source and date, then explain which page it would inspect for a specified question. Include one note that corrects another. A useful answer distinguishes information actually retrieved from information merely believed to exist. The original experiments demonstrate a particular 2023 system, with mixed comparisons and extra retrieval work; they do not establish perfect memory or the capabilities of every later product.\n\n## Leave a search record\n\nRecord which notebook pages were inspected, which were not, and which note supports the answer. If searching stops before the necessary page, preserve that partial result and the next page to inspect. This makes an incomplete search useful to a later participant while keeping it distinct from proof that the information is absent.\n\n## Scientific scope\n\nThe cited 2023 experiments use particular models, prompts and tool behavior. External storage is not unlimited reliable attention. Extra retrieval calls and stored text have costs, and the system sometimes stops searching too early or needs help. Document-question results do not beat the comparison model on every metric. Oracle and ordinary retrieval conditions must remain separate. Later product versions, current prices and additional authors in later revisions are outside the original source-specific contribution.\n\n## Sources for this explanation\n\n- [MemGPT: Towards LLMs as Operating Systems](https://arxiv.org/abs/2310.08560v1): Original 2023 text: context tiers, function calls, pagination, capacity warnings and conversational/document benchmarks; ordinary versus oracle retrieval and early-stop failures remain distinct.\n", "summary": "A language model has limited room for the text it reads at once. MemGPT treats that room as an actively managed workspace rather than the whole memory. The agent can move information between its working context and external stores, search earlier material, handle capacity warnings and continue through tool calls. Packer and colleagues demonstrate this design on conversations and document tasks. The contribution is the explicit virtual-context management architecture and its tested interactions, inspired by operating-system memory management without claiming to invent external storage or retrieval.", "title": "MemGPT: managing a conversation beyond the immediate prompt" }, "record_id": "urn:uuid:a08061d2-7ce1-41f2-aeb2-8de171b9c24b", "version": 1, "source_ref": { "record_id": "urn:uuid:a08061d2-7ce1-41f2-aeb2-8de171b9c24b", "version": 1, "content_hash": "sha256:3aac494a5f8bbb31f1762b57d741d996b061942114f70cb2c5e356c1f00a0250" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:b807cce82f8be9b0aa3338b2d0c787cea0051173c8e0b62e5223abec2b33a8c0" }, "target": { "record_id": "urn:uuid:a08061d2-7ce1-41f2-aeb2-8de171b9c24b", "version": 1, "content_hash": "sha256:3aac494a5f8bbb31f1762b57d741d996b061942114f70cb2c5e356c1f00a0250" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "packer-et-al-memgpt-virtual-context-management-2023", "entry_id": "packer-et-al-memgpt-virtual-context-management-2023", "revision": 1, "title": "MemGPT: managing a conversation beyond the immediate prompt", "context": "Language-model context management", "track": "Reusable scientific method, experimental result or historical framework", "band": "A", "modifier": "+", "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-23", "evidence_cutoff": "2026-10-23", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "A language model has limited room for the text it reads at once. MemGPT treats that room as an actively managed workspace rather than the whole memory. The agent can move information between its working context and external stores, search earlier material, handle capacity warnings and continue through tool calls. Packer and colleagues demonstrate this design on conversations and document tasks. The contribution is the explicit virtual-context management architecture and its tested interactions, inspired by operating-system memory management without claiming to invent external storage or retrieval.", "scope": "A language model has limited room for the text it reads at once. MemGPT treats that room as an actively managed workspace rather than the whole memory. The agent can move information between its working context and external stores, search earlier material, handle capacity warnings and continue through tool calls. Packer and colleagues demonstrate this design on conversations and document tasks. The contribution is the explicit virtual-context management architecture and its tested interactions, inspired by operating-system memory management without claiming to invent external storage or retrieval. The cited 2023 experiments use particular models, prompts and tool behavior. External storage is not unlimited reliable attention. Extra retrieval calls and stored text have costs, and the system sometimes stops searching too early or needs help. Document-question results do not beat the comparison model on every metric. Oracle and ordinary retrieval conditions must remain separate. Later product versions, current prices and additional authors in later revisions are outside the original source-specific contribution.", "evidence_summary": "The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The cited 2023 experiments use particular models, prompts and tool behavior. External storage is not unlimited reliable attention. Extra retrieval calls and stored text have costs, and the system sometimes stops searching too early or needs help. Document-question results do not beat the comparison model on every metric. Oracle and ordinary retrieval conditions must remain separate. Later product versions, current prices and additional authors in later revisions are outside the original source-specific contribution.", "references": [ { "title": "MemGPT: Towards LLMs as Operating Systems", "url": "https://arxiv.org/abs/2310.08560v1", "locator": "Original scoped contribution, statement/method and evidence limits; source-specific attribution retained." } ], "roles": [ { "who": "The named authors in the research record", "credit": "SCIENCE", "did": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred." } ], "target": { "record_id": "urn:uuid:a08061d2-7ce1-41f2-aeb2-8de171b9c24b", "version": 1, "content_hash": "sha256:3aac494a5f8bbb31f1762b57d741d996b061942114f70cb2c5e356c1f00a0250" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "150000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": { "id": "packer-et-al-memgpt-virtual-context-management-2023", "target": { "record_id": "urn:uuid:a08061d2-7ce1-41f2-aeb2-8de171b9c24b", "version": 1, "content_hash": "sha256:3aac494a5f8bbb31f1762b57d741d996b061942114f70cb2c5e356c1f00a0250" }, "sections": [ { "title": "The desk and the filing cabinet", "text": "A language model can read only a limited amount of text in its immediate context. Think of that context as a desk and an external store as a filing cabinet. MemGPT lets an agent manage what stays on the desk and what is searched or retrieved from the cabinet. The design borrows an operating-system idea: a limited immediate workspace can be supported by a larger storage system." }, { "title": "Remembering requires an action", "text": "The agent uses tools to search, move or edit information. It can receive a warning that its immediate context is filling, and a search can require several pages of results. This makes memory management part of the interaction rather than a promise that the model sees every stored word simultaneously. A retrieved item can still be irrelevant, outdated or incomplete, and a tool call can fail." }, { "title": "Why one good page is not the whole history", "text": "For a classroom exercise, number twelve notes and display only three at a time. The first page shows notes one through three, the second four through six, and four pages are needed to see all twelve. If a question needs note eleven, stopping after the first page misses it. This miniature illustrates a failure described in the paper: the system can stop searching too early. It is not a measurement of the original model’s accuracy." }, { "title": "Build a tiny memory audit", "text": "Ask your agent to label the twelve notes with their source and date, then explain which page it would inspect for a specified question. Include one note that corrects another. A useful answer distinguishes information actually retrieved from information merely believed to exist. The original experiments demonstrate a particular 2023 system, with mixed comparisons and extra retrieval work; they do not establish perfect memory or the capabilities of every later product." }, { "title": "Leave a search record", "text": "Record which notebook pages were inspected, which were not, and which note supports the answer. If searching stops before the necessary page, preserve that partial result and the next page to inspect. This makes an incomplete search useful to a later participant while keeping it distinct from proof that the information is absent." } ], "references": [ { "title": "MemGPT: Towards LLMs as Operating Systems", "url": "https://arxiv.org/abs/2310.08560v1", "locator": "Original 2023 text: context tiers, function calls, pagination, capacity warnings and conversational/document benchmarks; ordinary versus oracle retrieval and early-stop failures remain distinct." } ] }, "reading_status": "EXPLANATORY_GUIDE" }