Blog
hedwigaihedwigaihedwigai

Agents

Where do AI agents save their work? 24 artifact stores compared

An agent makes things. It writes a report, fills a table, builds a ledger or saves a file. The next session has to find that work and change it without breaking it. This post is about where that work lives. It is not about memory, the facts and preferences an agent remembers about a person.

We read the docs of 23 places an agent can save its work on 16 November 2026, and set our own beside them. For each one we asked four questions. Does it keep old versions? Does it refuse a write made from an old copy? Does it check what is written? Can another agent or a person read the same thing?

5 of 23

keep earlier versions of what the agent saved

2 of 23

refuse a write made from an out-of-date copy

3 of 23

check what is written against a type or a schema

The 24 stores

Each row says only what the maker’s own page says. “Not documented” means the page we read does not say either way. Our own store, hedwigai workbooks, is the last row.

24 places an agent can save its work, read from each maker’s docs on 16 November 2026. Each store name links to the page we read.
StoreMakerWhat the agent savesOld versions keptStale write refusedChecked on writeWho else can read itAccess
Agent BucketTencent CloudFiles in a Space, one per agent, user or teamNot documentedNot documentedNoShared Spaces and share linksConsole, REST API, COS SDK
ADK ArtifactServiceGoogleNamed files with a MIME typeYes, a number for every saveNot documentedNoThe same user's other sessions, with the user: prefixADK SDK
AgentCore session storageAWSA directory at a mount path, reset after 14 idle daysNoNoNoNo: each session sees only its ownRuntime configuration
Files APIAnthropicUploads, and files made by code execution and skillsNo: a file cannot change, so a new one replaces itNot needed: files cannot changeNoAnyone with a key to the workspaceAPI
Managed Agents session outputsAnthropicFiles the agent writes to its outputs folderNoNot documentedNoThrough the Files APIAPI
Managed Agents memory stores (beta)AnthropicText documents at pathsYes, each change kept 30 daysYes, by a content hashNoOther sessions, read-only or read-writeAPI, a mounted folder
Containers (Code Interpreter)OpenAIFiles in a container that expires after 20 idle minutesNot documentedNot documentedNoNot documentedAPI
Conversations APIOpenAIConversation items, tool outputs among themNoNot documentedNoNot documentedAPI
Agents SDK SessionsOpenAIChat history in SQLite, Redis and other databasesBranching in one session typeNot documentedNoNot documentedSDK
Foundry Agent ServiceMicrosoftFiles in your Azure Storage, threads in Cosmos DBNot documentedNot documentedNoInside your own tenantAPI, SDK
LangGraph checkpointersLangChainThe graph's state, per threadYes, the checkpoint historyNot documentedNot documentedThe same threadSDK
LangGraph StoreLangChainKey-value items in namespacesNot documentedNot documentedNot documentedOther threadsSDK
Flows @persistCrewAIA flow's state in SQLiteSnapshots, to resume or forkNot documentedYes, Pydantic types on the stateNot documentedSDK
Agents SDK stateCloudflareJSON state and SQL tables in a Durable ObjectNot documentedNot documentedYes, a validateStateChange hookConnected clients, kept in syncSDK
Artifacts (beta)CloudflareGit repositories of filesYes: commits, branches and forksNot documentedNoAny Git clientWorkers binding, REST, Git
Pause, resume and snapshotsE2BThe whole sandbox, files and memorySnapshots onlyNoNoNot documentedSDK
VolumesModalFiles, seen by others after a commit and reloadNoNo: the last write winsNoMany functions and sandboxesSDK, a mount
VolumesDaytonaFiles on storage backed by S3NoNo: the last write winsNoMany sandboxesSDK, a mount
Sandbox persistence and DrivesVercelThe sandbox's file systemSnapshots onlyNot documentedNoAgents in the same Linux groupSDK, CLI
AgentFSTursoFiles, state and history in one SQLite fileSnapshots and an audit logNot documentedNoWhoever has the fileCLI, SDK
Notion MCPNotionPages and databasesNot documentedNot documentedNot documentedPeople in the workspaceMCP
Supabase MCPSupabasePostgres tablesDatabase branching, experimentalNot documentedYes, the Postgres schemaAnyone with database accessMCP
Contents API and GitHub MCPGitHubFiles in a repositoryYes, Git historyYes: an update sends the replaced file's shaNoAnyone with access to the repositoryREST, MCP
WorkbookshedwigaiTables and documents as markdown tabs, or an Excel, Word or Drive fileYes, every write; any version can be restoredYes: a write from an older version is refusedYes, compiled against the workbook's schemaOther agents, and people in the web appMCP, web app

Tencent's Agent Bucket: one Space per agent

Tencent Cloud launched Agent Bucket in July 2026. It is a bucket type inside its object storage, COS. A bucket is split into Spaces. Each Space belongs to one agent, one user or one team, and has its own quota, permissions and access token. Agents save reports, images, code and intermediate results there, next to the files users upload. Tencent says its own agent products run on it.

Agent Bucket solves who owns a file and who may read it. Its docs do not describe versions of a single file, a check on what is written, or a refusal of a write from an old copy. A file is bytes at a path.

Few stores keep old versions

Only 5 of the 23 keep earlier versions of a single thing the agent saved: Google’s ADK artifacts, Anthropic’s memory stores, LangGraph’s checkpoints, Cloudflare’s Artifacts and GitHub. The sandboxes keep snapshots of a whole disk at most. Anthropic’s Files API keeps no versions because a file can never change. A changed report is a new file, and a deleted one cannot be recovered.

Fewer refuse a stale write

Two agents, or one agent in two sessions, often work from copies. When the older copy is written back, it erases the newer one. Only 2 of the 23 stop this. Anthropic’s memory stores take a hash of the content the write was based on, and GitHub’s contents API takes the sha of the file being replaced. Modal and Daytona say in their docs that the last write wins.

Almost none check what is written

A file store accepts any bytes. Only 3 of the 23 can check a write against a shape: CrewAI through Pydantic types on a flow’s state, Cloudflare through a hook that can reject a change, and Supabase through the Postgres schema. None of them keeps versions, refuses stale writes and checks each write together.

Where hedwigai workbooks sit

hedwigai is a system of record for agents, reached over MCP. An agent keeps its work in workbooks. A workbook is a set of tabs, each a table or a page of notes, or an Excel or Word file. It can also be linked to a file in Google Drive or Dropbox, and that file stays the record.

Every write makes a new version, and any version can be read or restored. A write says which version it was based on, and it is refused if someone wrote a newer one since. Each write is compiled against the workbook’s schema: its columns, types, required cells, keys and references. A bad cell comes back as an error, the way a compiler answers, and the agent fixes it and writes again. Claude, ChatGPT, Codex and Cursor read the same workbook, and so do people in the web app.

claude mcp add --transport http --scope user hedwigai https://mcp.hedwigai.com

In Claude, ChatGPT or Cursor, add https://mcp.hedwigai.com as a custom connector. The server is in the official MCP registry as com.hedwigai/hedwigai. Sign-in is OAuth, so there is no token to paste.

How we read them

We read each maker’s own documentation, linked from the store’s name in the table. A cell says “Yes” only where the page says so. We left out stores whose only description was a press report or a code sample: Vertex AI Agent Engine’s code-execution files, ByteDance’s Volcengine AgentKit, Alibaba’s AgentRun, Archil and Mesa. Memory services such as Mem0, Zep and Letta are left out on purpose, because they store what an agent remembers rather than what it makes.

More head-to-head evals

Particle-1, DeepSeek V4.1 Flash and GLM-5.3-Flash in Claude Code with hedwigai workbooksParticle-1, DeepSeek V4.1 Flash and GLM-5.3-Flash, 180 runs graded blind, rows right: 91.2%, 89.4% and 90.8%, intervals overlapping.
MeasureParticle-1DeepSeek V4.1 FlashGLM-5.3-Flash
Rows right, blind91.2%89.4%90.8%
Median seconds per run375315144
Read the comparison
GLM-5.3 against DeepSeek V4.1 Flash on 75 questionsTwo models answer the same 75 questions; one is always right, the other 14× cheaper.
MeasureGLM-5.3DeepSeek V4.1 Flash
Right, of 757571
Seconds a question, median3.91.8
Read the comparison
How we made a long-running AI agent for knowledge work 4× fasterOne cited knowledge work task through 13 experiments: 937 s to 233 s, with the model held constant.
MeasureFirst runBest since
Landscape, seconds937233
Memo, seconds759602
Read the report
Ten models plan a change to a spreadsheet: which do it in under a minute, traced call by callTen models drive one agent harness through the same spreadsheet request; six write the plan, timed and traced call by call.
MeasureQwen3 235BGLM-5.2Gemini 3.8 FlashClaude Opus 5
Seconds to the plan, median27485794
Plans written, of three runs2333
Read the comparison
SubQ, jev, Perplexity's Decider and Bespoke's Nimble on biotech diligence memos: finding the sentences a column is aboutFour decision models judge every sentence of ten biotech diligence memos for five columns, scored against sentences a person marked.
MeasureSubQjevDeciderNimble
Found, mean of five columns at section length64%65%66%68%
Mechanism of action, found at section length57%51%61%54%
Read the comparison
Leads with a reason to write: hedwigai ask, Apollo and treg on one brief40 leads each, every one checked: the right person, an inbox that accepts, and a dated reason to write.
MeasureaskApollotreg
Usable leads13 of 407 of 406 of 28
Cost per usable lead$0.037$0.27$0.23
Read the comparison
Self-hosted decisions for ask: Perplexity's open Decider in place of jev on 92 company questionsThe decisions inside hedwigai ask made by Perplexity's open Decider on two NVIDIA RTX 3090s, against jev.
Measureask with jevask with Decider
Questions right, of 928989
Wrong answers12
Read the comparison
Clinical decisions: jev and ask on System One, PubMedQA, SciFact and a 370-decision setEdgeEvals' System One board, PubMedQA, SciFact, and a clinical set built to System One's rules.
Measurejevask
Our 370 clinical decisions, right94.6%93.0%
PubMedQA, right (best EdgeEvals scored: 62.2%)77.0%74.4%
Read the comparison
Executive lookup for US-listed companies: Claude Opus 5.5 and ask on 95 questionsWho runs a public company, and how to email them: a frontier model from memory against hedwigai ask.
MeasureaskClaude Opus 5.5
Questions right, of 955963
Small companies right, of 20139
Read the comparison
Rental comps: floor-plan rents and amenities on 14 listings, five assistants comparedRent per floor plan and amenities for US rentals: hedwigai realestate-us against Claude, GPT, Gemini and Perplexity with web search.
Measurerealestate-usBest of four assistants
Floor-plan rents right, of 262623
Amenities claimed the listing doesn't state12129
Read the comparison
Company screening from SEC filings: jev, Vela 2.0 and ask on 92 yes/no questionsYes-or-no questions about public companies, answered by two decision models alone and by hedwigai ask.
Measureaskjev
Right, of the questions answered98%79%
Wrong answers, of 92219
Read the comparison
Defect localization in industrial images: four frontier models and a trained localizerFour frontier models asked to put a box around what they just described.
MeasureTrained localiserBest of four frontier models
Mean overlap with the defect (IoU, 0 to 1)0.4210.109
Read the comparison