Local first · published CLI 0.2.9
Your first useful memory.
Initialize local memory, then connect your coding agent. No account or API key required.
1. Start in your project
In your project terminal (Node 18+ and Python 3.10+):
npx @agentevo-cli/agentevo@0.2.9 initSelect this command to copy manually, or use Copy.
What happens: the pinned launcher fetches the CLI and creates your project identity and local vault. Keep its key in your OS keychain. Nothing is uploaded. An empty memory is normal; initialization does not connect your coding agent or learn a lesson yet.
Python only? Install with uv
Use uv with Python 3.10+. This isolates the CLI from your project dependencies. You can use bare agentevo in place of the pinned npm launcher in the examples below.
Select these commands to copy manually, or use Copy.
uv tool install agentevo-sdk==0.2.9
agentevo initHeadless setup, versions and diagnostics
Headless use needs one retained vault key through your secret manager. Never create a new key each time. Choose agentevo-sdk on PyPI or @agentevo-cli/agentevo on npm; unscoped names are other products. This guide uses published SDK/npm 0.2.9 (local 0.1.9, core 0.1.8). For diagnostics, run npx @agentevo-cli/agentevo@0.2.9 doctor. Doctor checks local configuration and does not attest host interception.
2. Connect your coding agent
Let your agent handle the setup
Paste this request into your coding agent in the same project. It can run the setup commands and explain any proposed connection before you approve it. You can skip the terminal step above: the request includes initialization.
Copy the request and paste it into your coding agent.
Read the setup request
Set up AgentEvo in this repository using the published 0.2.9 CLI. Run npx @agentevo-cli/agentevo@0.2.9 init from this project. Keep memory local; do not sign in, enable sync, upload data, run verification replay or promote objects during setup. Do not ask me for API keys. Ask which coding agent I use. For Claude Code 2.1.295+, run npx @agentevo-cli/agentevo@0.2.9 host setup claude-code to generate a preview. Show me the proposed files, hooks and limits. Ask me to approve that exact preview before applying it with --apply-preview-sha and the returned preview_sha256. Re-preview if anything changes. Then check host status. Do not grant new host permissions or create a test context automatically. For other coding agents, offer the project-local bundled guides with npx @agentevo-cli/agentevo@0.2.9 skills install --host agents. Explain the proposed files and get my approval first. Explain that Beta transcript imports are explicit and automatic native capture is not configured. Summarize what is initialized and what is actually connected. If Claude Code was connected, tell me to restart it. Next, suggest a small tested task. Explain that an empty memory is expected and local setup alone is not a learned lesson or proof of host enforcement.
Prefer to connect Claude Code manually?
Initialization creates local memory. Connecting a host is a separate reviewed action. Supported Claude Code integration requires Claude Code 2.1.295+ and a reviewed setup preview.
Select these commands to copy manually, or use Copy.
npx @agentevo-cli/agentevo@0.2.9 host setup claude-codeRead the proposed files, hooks and limits. Apply only the exact command and preview digest printed by the CLI when you accept that configuration. Re-preview after changes. Setup compiles Candidates and grants no host permissions. Selected test context, workflow execution and high-risk approval have additional policy/evidence requirements.
Use agentevo doctor to inspect configuration; actual host callbacks and a controlled host-action check are separate evidence. For a strict CI gate, review agentevo ci --help and prepare a nonempty pinned manifest.
Learn from a completed session
Show the manual learning steps
Start with a completed Claude Code task in this project and working project tests. No session yet? Initialization still works: finish a small tested task, then return here. You can also explore the illustrative lifecycle first.
Select these commands to copy manually, or use Copy.
npx @agentevo-cli/agentevo@0.2.9 capture --latest
npx @agentevo-cli/agentevo@0.2.9 compile --latest
npx @agentevo-cli/agentevo@0.2.9 verify --latest --local --workdir .
npx @agentevo-cli/agentevo@0.2.9 promote --latest --to Verified --local
npx @agentevo-cli/agentevo@0.2.9 recall "fix flaky test" --capabilities Bash,Edit,Read
npx @agentevo-cli/agentevo@0.2.9 reportReplace the recall query with your task. Expected: capture stores events locally, compile creates Candidates, verification reports evidence, and eligible objects promote to Verified. Recall can return matching local objects; the report opens from agentevo-report.html. Empty recall is a valid result, not a reason to invent a match.
Review the declared commands before verification: replay runs tests in your working copy and is not sandboxed. Inspect current_replay, integrity and task correctness separately. Interactive exit zero can contain failed or inconclusive evidence; a Verified object is not independently proven correct. Promotion refuses unmet gates.
I use Codex, Cursor or OpenCode
These Beta adapters use an explicit transcript in the supported interchange format. They do not automatically read every native agent history store. Start with agentevo capture --help; select your adapter and reviewed file explicitly.
Select these commands to copy manually, or use Copy.
npx @agentevo-cli/agentevo@0.2.9 capture --adapter codex ./session.jsonlsession.jsonl is your supported export, not a file created by this guide. Review it for sensitive content first.
Optional: connect cloud features
Set up searchable cloud sync
Local recall in an initialized project works without login or upload. The console displays authenticated cloud activity. To enable searchable sync, review the Terms and Privacy Policy, create an appropriate API key in the console, then run these commands only if you intend that consent and upload:
Select these commands to copy manually, or use Copy.
npx @agentevo-cli/agentevo@0.2.9 login
npx @agentevo-cli/agentevo@0.2.9 sync mode --set searchable --terms-version 2026-09-21 --yes
npx @agentevo-cli/agentevo@0.2.9 sync push
npx @agentevo-cli/agentevo@0.2.9 recall "fix flaky test" --capabilities Bash,Edit,Read --cloudLogin checks the hosted API before storing credentials in the OS keychain. Searchable sync sends minimized signed objects; explicit cloud recall sends its query and selected sanitized context. Optional BYOK embeddings contact the selected model provider and can incur its charges. Read the data boundary before enabling either.
Objects pulled to another machine arrive as Candidates. Verify them in that machine's project before promotion and local recall. Offline peers learn cloud revocation changes on later receipt/sync.
When something needs attention
The command is missing or reports another product
Check the exact package name and agentevo --version. Open a new terminal after tool installation, or use the pinned npm launcher. For uv PATH instructions, run uv tool update-shell and review its output.
Doctor reports a missing key or unconfigured host
Retain the original OS-keychain key. For headless use configure one persistent vault key through private secret management; a new random key on every run cannot open existing memory. An unconfigured host is separate from local readiness: follow the reviewed preview above if you want integration.
No session, no candidate or no recall hit
Run from the initialized project where the session occurred. Claude latest selection is project-bound. For Beta adapters provide a supported export explicitly. Compile does not manufacture reusable evidence from every run. Recall withholds ineligible objects and reports hints; inspect those before verifying/promoting the exact object.
Verification is inconclusive or promotion refuses
Check the JSON and supported test command, dependencies and --workdir .. Review and run your project tests first, then reverify. Do not treat interactive exit zero as a strict CI pass or force a trust state to remove the warning.
Cloud sign-in, quota or connectivity fails
Initialized local recall keeps working. Explicit cloud requests can fail or pause at quota; inspect their error/reset guidance. Export, deletion and containment have their documented exceptions. Review the workspace sync mode before retrying upload, and check current state before repeating a timed-out mutation.
Optional: avoid regenerating a working implementation
For agent-host builders using SDK 0.2.9
Start with the working files you already keep. For eligible historical work, SDK 0.2.9 adds experience.read and agentevo_sdk.reuse.try_reuse. The host supplies its existing approved current-contract sandbox validator. At most three candidates are validated; a passing implementation is rechecked for eligibility before the host stages it through normal permissions and skips generation. A miss or failed test continues normal task work.
Installing the CLI alone does not enable this orchestration. The trusted callback binds the current task/tests/environment to a contract SHA256 and actual validation receipt. Cleanup failures stop the host. A partial Edit remains a fragment; a passing old lesson alone does not grant action authority.
An explicit read returns bounded untrusted C3 code, which your coding host may forward to its model provider. Review that handling and current tests before use. A memory hit alone is not a saved request. The complete repeat-work study includes paid cold acquisition, failures and the plain file-cache control; it does not establish customer net ROI or human time saved.
Select this command to copy manually, or use Copy.
npx @agentevo-cli/agentevo@0.2.9 experience read --helpThe published SDK README describes the host API and its limits. This optional integration adds no prerequisite to the local initialization above.
Understand what the result means
Raw captures are local by default. Pattern-based secret detection is finite; review sensitive material before sync. Recall is advisory data, verification is evidence, promotion controls eligibility, and supported host policy governs actions. None of these steps trains model weights or guarantees a correct task.
Activity counts, recall hits and estimated avoided calls are separate from money saved. We have not established customer time savings or ROI. Read the measured outcomes and dated engineering evidence.