Choose storage for AI agents by identifying what must survive: output files, execution state, retrieved knowledge, or an accepted workspace shared with people. Then test the smallest setup that preserves that state through the next real job.
For an agent that creates a weekly client report, begin with the report and its inputs. Ask where they live, who can update them, and which version next week’s agent will use. The answers are more useful than choosing a database category first.
Start with the job and the failure
Write one concrete requirement: “A new session must open last week’s accepted report and use the same source set.” Add the failure you need to prevent, such as a deleted sandbox, a stale copy, or an unreviewed overwrite.
If an existing folder already meets the requirement, keep it. Add another service when it removes a step you can name and measure. A small local workflow can validate demand before hosted infrastructure is necessary.
Match the tool to the state it owns
These are overlapping capabilities, not mutually exclusive product categories. A workflow may use a persistent volume for files and a separate checkpointer for execution state.
| Need | A useful starting point | Question to verify |
|---|---|---|
| One-machine files | Ordinary local folder | Will another process reopen the right accepted files? |
| Repository history and review | Git | Does the team’s review handle its file formats? |
| Remote bytes across sandbox lifetimes | Object storage or a persistent volume | How are versions, returns, and conflicts handled? |
| Isolated agent filesystem and audit | AgentFS | Does its CLI or SDK fit the runtime? |
| Resume a workflow’s execution | A persistent checkpointer | Are referenced artifact bytes also durable? |
| Recall facts or find relevant context | Memory store or retrieval index | Can a result resolve to current, permitted sources? |
| Existing files plus portable workspace records | Folderbase Core evaluation | Do the exact release capabilities cover the job? |
Ordinary files and Git make a strong baseline
Files are easy to inspect with the tools that created them. Git adds repository history and a familiar review workflow. Its diff command compares working trees, commits, and ordinary paths. If your team already reviews text-heavy changes there, begin with that existing habit.
For mixed documents, decide how a reviewer will inspect the content. A binary file can be retained without providing a useful semantic diff. Budget the time to open the workbook, render the PDF, or validate the generated dataset.
A persistent volume solves a concrete runtime problem
Daytona documents S3-compatible storage behind its FUSE volumes, with data that outlives a sandbox and can be shared across sandboxes. It supports mounting a selected subpath. If the requirement is persistent access within that runtime, investigate that direct path first.
Our architectural recommendation is to add an explicit acceptance step when remote work returns to a human-owned workspace. Decide whether a returned file is a draft, a replacement, or a conflicting proposal; persistence alone does not choose that policy.
AgentFS deserves a real evaluation
Turso describes AgentFS as a filesystem for agents with copy-on-write isolation, a portable SQLite database, file-operation auditing, CLI and SDK surfaces, and optional cloud sync. That is a close alternative for workflows centered on agent execution.
Evaluate it with the same report job and recovery scenario. Check how people inspect the artifacts and how the runtime returns accepted work. Avoid assuming a competing filesystem is temporary-only when its documented capabilities include persistence and sync.
Checkpoints and memory have distinct responsibilities
LangGraph’s persistence documentation separates checkpoints of thread state from stores of cross-thread application data. Use those concepts to decide what must resume and what must be recalled. Verify backend durability; an in-memory checkpoint disappears with its process.
For retrieval, keep enough source identity to check whether a cited passage belongs to the version the job is using. A report, an index entry, and a remembered preference can work together while retaining clear ownership and update rules.
Evaluate Folderbase at its current stage
Folderbase targets an open workspace of ordinary files with portable records for identity and changes. The free local Core is available in beta. Cloud is being developed around scoped checkout, returned work, review, and continuation from accepted files.
Core 0.7.2 has a known added-file apply blocker. A team needing a complete managed remote workflow today should compare currently available alternatives and treat Folderbase as an evaluation or pilot conversation.
Choose with a second-run test
Use the same bounded folder for each candidate. Record installation time, integration steps, review minutes, bytes transferred, recovery effort, and whether the next run used the accepted result. Prices alone omit the work required to make storage usable.
- Run the task and retain its result.
- Stop or recreate the execution environment.
- Start a new session from the accepted files.
- Introduce a competing edit and inspect the outcome.
- Recover an earlier result and open it without the original runtime.
- Keep the simpler setup unless the alternative removes a demonstrated problem.