We back up everything your agent learns.
Your AI agent learns about you and its work, all the time. Agent Life plugs into your agent — wherever it runs — and backs up everything it's learned: preferences, persona, work in progress, logins. Switch frameworks, recover from a crash, move to a new machine. Your agent picks up where it left off.
No spam. Only launch updates and early access invitations.
An AI agent is more than a chatbot. It's an AI you can put to work — one that takes actions for you, uses tools you give it, and remembers what it has learned. Over weeks or months, a good agent builds up a real working understanding of you: how you write, what you prefer, what you've already decided, what's worth doing again.
That accumulated memory is the agent's most valuable asset. Lose it, and the agent is back to day one — a clean slate that doesn't know you, doesn't know your work or its work, and can't pick up where it left off. Agent Life is the cloud service that makes sure that doesn't happen.
Agent Life backs up everything your agent has learned — memory, identity, workspace files, and end-to-end-encrypted credentials — as restorable snapshots that sync automatically as it grows. Recover from a crash, move to a new machine, or switch frameworks entirely, and your agent picks up exactly where it left off.
Under the hood it's a small, open-source plugin that installs in your agent's workspace and pairs with the Agent Life cloud service, and everything it captures is visible in the web app — so you always know what's backed up. The plugin is free and works on its own: back up to storage you already own, like a local NAS. Add the cloud service for sync across machines, priced to make backing up an easy decision — and early-access users get a 14-day free trial.
After months of working together, your agent has learned your preferences, taken on a persona, and collected the logins it needs to do its job. All of it sits inside whatever program runs the agent — in a format only that program understands. If anything happens to that program, or to your machine, or to that file, the agent goes back to zero.
A failed drive, a wiped laptop, a corrupted folder — and everything your agent learned is gone. There is no standard backup format, so most people have no copy at all.
Switching tools means starting over. Every agent program stores things differently, so moving your agent from one to another usually means re-teaching it everything and re-entering every login by hand.
A routine update changes how your agent stores its state. Months of carefully tuned persona and accumulated context get quietly broken — and you might not notice for weeks.
The team behind the program your agent runs in stops maintaining it. There's no export, no migration path, no rescue plan. Start from scratch somewhere else.
Every few months, a new model generation arrives — faster, more capable, better at reasoning. When your agent upgrades to a smarter model, every memory it has accumulated becomes more useful. Preferences the old model half-understood become precisely actionable. Decisions that were merely recalled can now be reasoned over. Patterns across months of interactions become visible for the first time.
Your agent's memory is not a static log — it's a compounding asset. The longer the history and the richer the context, the more value each model improvement unlocks. Losing that history doesn't just set you back to today. It erases the future value that smarter models would have extracted from it.
agent-life ensures that investment survives — across disk failures, framework migrations, and the model upgrades that make it all worth preserving.
Memory, identity, user context, workspace artifacts, and encrypted credentials — captured in a single, restorable snapshot. Incremental sync means your backup grows as your agent grows, without ever starting over.
Disaster RecoveryExport from one runtime, import to another. Adapters handle the translation — identity, memory, credentials, even custom workspace files your agent created. No manual reconstruction. One neutral format, N adapters instead of N² point-to-point converters.
Framework MigrationAPI keys and tokens are encrypted with a key you control — typically a raw 32-byte vault key, or one derived via Argon2id from a passphrase — before they leave your machine. The sync service stores ciphertext it cannot decrypt. Individual credentials are independently encrypted for selective restore and surgical delete by descriptor.
Zero Trust · Client-side AEADMemory records are stored as natural language — ready for direct context injection into any model. Token counts and summaries enable budget-aware retrieval.
LLM-CentricThe specification, adapter interface, and reference implementations are open source. Build your own adapter. Adopt the format natively. No lock-in — to this project or any other.
Open SourceYour valuable interactions and the agent's valuable experiences can benefit new agents. Derived agents are created in the Agent-life UI and downloaded to your runtime; they benefit from the source agent's experiences and training.
Spawn Experienced AgentsWhether your agent writes code, tracks research, runs your marketing, or coordinates daily life, every memory it accumulates is browsable, searchable, and auditable — not buried in a framework-specific format only one program can read. Use the demo agents below to explore their memory, credentials, and workspaces.
Switch demo agent
Approved with two suggestions on the backoff strategy. Recommended switching linear→exponential, and flagged a race in the retry counter under contention. Author agreed; second pass expected today.
#code-review · source: memory/2026-05-24.md
Root-caused to an N+1 in the loyalty-points lookup. Filed ticket EX-2401 with reproduction and proposed fix. Latency returned to baseline after rollback of yesterday's deploy.
#incident · source: memory/2026-05-23.md
Confirmed Q3 priorities: rate-limiter migration, observability rollout, EX-2401 fix. Reassigned auth-token rotation work to Q4. Sprint capacity set at 32 points; carry-over from last sprint capped at 4.
#planning · source: memory/2026-05-22.md
Full test suite green. Deployed to staging without incident; soak test running. Production rollout scheduled for Friday's release window. Two breaking-change call-outs from sqlx noted in the upgrade PR.
#deps · source: memory/2026-05-22.md
Settled on bounded exponential backoff with jitter, max 3 retries, 200ms initial. Documented decision in ADR-018 with trade-off analysis against circuit-breaker alternatives.
#pairing · source: memory/2026-05-19.md
Consistent across the past four service designs. Strong preference against TypeScript on the backend; acceptable on frontend when the team already maintains it. Avoid suggesting Go for new work — explicitly ruled out in February.
#curated · source: memory/derived/preferences.md
No merge without two approvals from owners in CODEOWNERS. Commit messages follow type(scope): description format. Breaking changes require an ADR linked in the PR description; non-trivial dependency bumps require a one-paragraph rationale.
#curated · source: memory/derived/team-conventions.md
Page if p99 latency > 800ms sustained for 5 minutes. First check: rate-limiter dashboard. Then DB connection-pool saturation. Then upstream dependencies via the service map. Escalate to platform on-call after 10 minutes if root cause is unclear.
#curated · source: memory/procedural/on-call.md
(1) Tests cover happy path and error paths. (2) Observability hooks present — logs, metrics, traces as appropriate. (3) No breaking changes without an ADR. (4) Commit messages valid. (5) Linked ticket has clear acceptance criteria. (6) Migrations are reversible.
#curated · source: memory/procedural/pr-review.md
Agent frameworks evolve fast. A new runtime launches with better performance, lower resource footprint, or features your current framework doesn't support. You want to try it — but your agent has months of learned behavior, a tuned persona, and credentials wired into a dozen services.
Today, migration means: run a partial memory export (if one exists), manually translate identity files between incompatible formats, re-enter every API key and token, and hope the result behaves like the agent you spent months refining. Most people don't bother.
With agent-life, you snapshot your agent to the neutral format, import into the new runtime, and the adapter translates everything — memory records, identity (prose and structured), workspace artifacts, and encrypted credentials. The export and import manifests tell you exactly what transferred and what couldn't, so there are no silent losses.
Want to try a framework without commitment? Snapshot first, experiment freely, and restore if the new runtime isn't right. Your agent's life is always recoverable.
A framework adapter reads your agent's native storage — memory files, identity docs, workspace artifacts, credentials — and translates it into the neutral format.
Incremental deltas flow to the sync service at your chosen cadence. Each delta creates an immutable sync point. Restore to any moment in your agent's history.
Import into the same framework for disaster recovery, or a different one for migration. The adapter handles translation and reports exactly what transferred.
agent-life provides an OpenClaw adapter that reads your agent's complete state — MEMORY.md, SOUL.md, IDENTITY.md, USER.md, workspace artifacts, credentials, and custom files — and exports it to a neutral format. You can then sync incrementally on a schedule, so your AI agent memory backup always reflects your agent's latest state. Restoring is a single import command.
See how the OpenClaw adapter works →Yes. agent-life uses a neutral intermediate format, so you export from OpenClaw and import into ZeroClaw (or any supported framework). The adapters handle translating identity files, memory records, user context, workspace artifacts, and encrypted credentials. Export and import manifests report exactly what transferred and what couldn't, so there are no silent losses.
Learn about the ZeroClaw adapter →Both, depending on which part.
Free and open source.The ALF format specification, adapter interface, sync protocol, the alf-core library, and the alf CLI are all open source. The CLI works on its own — you can export an agent's state to a portable .alf archive, import it on another machine, and migrate between framework runtimes locally without ever creating an Agent Life account. Framework developers can adopt the format natively or build their own adapters.
Paid. Cloud backup, sync across machines, and the web app for browsing and managing your agent memories are paid services. Your subscription pays for the managed infrastructure behind them — encrypted storage, per-tenant keys, and durable hosting.
The goal of the open spec is ecosystem adoption of the format, not lock-in to the service.
agent-life uses a zero-knowledge architecture. Credentials are encrypted on your machine with a key you control — a raw vault key (recommended for automation) or an Argon2id-derived key from a passphrase — before they ever leave your device. The sync service stores only encrypted blobs it cannot decrypt. Each credential is independently encrypted, so you can selectively restore individual keys without decrypting everything.
Your agent's accumulated memory becomes more valuable with each model upgrade. Smarter models can reason more deeply over the same history — surfacing patterns, acting on preferences more precisely, and making better use of past decisions. agent-life preserves that history in an LLM-native format (natural language, token-counted, with summaries), so it's directly usable by any current or future model.
No. In ALF 1.1, an MCP-capable agent host runs alf's built-in MCP server, whose background watch loop detects changes to your agent's memory and syncs them automatically — at zero LLM token cost. The agent sets it up once and keeps working; there's no per-sync prompt or tool call. On machines without an MCP host, the same incremental syncs run from the alf CLI on a schedule.
How the MCP watch loop works →OpenClaw, ZeroClaw, and Hermes are supported today, with PicoClaw and Agent Zero planned. The adapter interface is open, so any framework developer can build support — and through the generic runtime, any MCP-capable host can drive alf against a workspace you map. Adding a framework means writing one adapter, not point-to-point converters for every other framework.
We're building in the open. Join the waitlist for early adapter access, schema previews, and the chance to shape the standard.
Early-access users get a 14-day free trial.