memorax-ai/memorax-code

_{Beyond code, it remembers how your architecture evolves and how your engineering unfolds.}

English · 简体中文

Coding agents are good at the task in front of them, but a new session often starts without the architecture, failed attempts, repository rules, or working preferences established before it.

MemoraX Code gives Codex, Claude Code, CodeBuddy CLI, WorkBuddy, DeepSeek Harness, OpenCode, Trae, and Cursor a shared memory layer for that context. It can recall prior engineering knowledge, capture reusable lessons from completed work, maintain repository knowledge, and carry your procedures and preferences into future sessions.

The goal is not to remember everything. It is to bring back the small amount of memory relevant to the current task so the agent can reach useful investigation and validation sooner.

Prepare Node.js 20+ (Node.js 24 LTS recommended) and at least one of Codex, Claude Code, CodeBuddy CLI, WorkBuddy, DeepSeek Harness, OpenCode, Trae, or Cursor.

For DeepSeek Harness (DSH), current releases require Node.js ^22.19.0 || >=24.0.0. Install or initialize DSH first, create at least one Profile, and ensure pnpm is on PATH before running setup. MemoraX Code does not install or update DSH.

On Linux, guest credentials require /usr/bin/secret-tool from libsecret and an available Secret Service in the current user session. For Remote SSH, WSL, or Dev Containers, install MemoraX Code in the same environment as the coding agent. MemoraX search and writeback require network access.

npm install -g @memorax/memorax-code

This installs the package; it does not start interactive setup. Do not use --ignore-scripts: npm lifecycle scripts safely stop and restore an existing running managed Backend during package replacement.

Create a MemoraX account or use an existing one, then run from a normal interactive terminal:

memorax-code setup --existing-account

Follow the setup prompts to enter your MemoraX username and API key locally.

For a coding agent without an interactive terminal, pass the API key through stdin. The examples assume MEMORAX_SETUP_API_KEY is already provided by the caller. Do not put the key in command arguments or project files.

printf '%s\n' "$MEMORAX_SETUP_API_KEY" | memorax-code setup --existing-account --non-interactive

In Windows PowerShell:

$env:MEMORAX_SETUP_API_KEY | memorax-code.cmd setup --existing-account --non-interactive

This explicit command replaces the saved key and uses the detected local username and system language. It reports API Key match: true after local configuration and readiness checks; this does not verify cloud credentials. See non-interactive setup for input and reuse behavior.

To start immediately and connect an account later, run:

memorax-code setup

Default setup reuses a complete existing connection. Otherwise, it detects your local username and language, asks when needed, and creates or restores guest credentials. To replace the saved connection, use memorax-code setup --reconfigure for guest mode or memorax-code setup --existing-account for a registered account.

To keep your guest memory when registering later, first run this command directly in your local terminal:

memorax-code account --show-mark-id

Both setup paths automatically detect supported coding agents. Restart or refresh every detected coding agent after setup.

Cursor installs independently of Claude Code under ~/.cursor/hooks.json and ~/.cursor/skills/memorax-code/; setup preserves its third-party integration setting. Use the Skill for CLI Search and manual Add. Automatic Add reads verified native database turns and requires Node.js 22.13+ with built-in SQLite (Node.js 24 recommended). It supports ordinary prompts, edited resends, and continuations bound to an observed preceding turn; interrupted or ambiguously correlated content is skipped. Prompt Hooks inject trusted User Profile preferences on the first eligible turn and Procedure Memory on the configured reminder cadence. The default Procedure cadence is turns 1, 6, and 11 (then every five turns). After a recorded compaction is verified against the native database, the next nonempty, registered prompt restores Profile and personal reminders; Procedure keeps its normal cadence. Missing evidence skips recovery, and restoration during the same continuing task is not guaranteed. Repo Memory initial builds and policy-based maintenance use Cursor's native background subagent; no separate Cursor CLI or CLI login is required. Cursor may request normal tool approvals. See Cursor configuration.

Open a project, start a new client session, and send one prompt. Then run these commands from the project directory:

memorax-code --version
memorax-code status
memorax-cli status

In Windows PowerShell, use memorax-cli.cmd status. A configured integration may still report hook-runtime=unverified until the client executes its Hook. After a Hook executes successfully, that client's Hook runtime should change to observed.

memorax-code status checks the local Backend and client integrations; memorax-cli status checks the local memory configuration and workspace scope. Neither command sends a test request to MemoraX. A real search or write verifies remote connectivity and credentials; follow the cross-session example below. For client-specific diagnostic commands, see Troubleshooting.

Package installation does not launch setup automatically; run one of the setup commands above, using stdin mode when an agent has no interactive terminal. For incomplete setup or unavailable memory, start with the status commands and follow Troubleshooting.

Both commands are included in the same package. Follow the Windows PATH repair steps to bootstrap setup or repair a stale terminal environment.

Clone the example repository from the product website, then open Codex, Claude Code, CodeBuddy CLI, WorkBuddy, DeepSeek Harness, OpenCode, Trae, or Cursor in the project directory:

git clone https://github.com/SWE-agent/test-repo.git
cd test-repo

Invoke the Skill as $memorax-code in Codex or /memorax-code in Claude Code or DeepSeek Harness. In OpenCode, CodeBuddy CLI, WorkBuddy, Trae, or Cursor, ask the agent to use the memorax-code skill by name. The prompts below use its product name and work in all supported clients.

Send these prompts in order in the same session:

1. Use the MemoraX Code skill to build Repo Memory, retrieving only the latest 3 issues, pull requests, and commits.
2. Review the recent Repo Memory issue: the zeroth number was once calculated incorrectly. Avoid repeating the same problem now.
3. Use the MemoraX Code skill to remember the engineering lesson from this coding task.

Close the current conversation, start a new session in the same repository, and send:

Use the MemoraX Code skill to recall the earlier engineering lesson and suggest what to check.

The agent should retrieve the saved lesson and use it to make suggestions for the current repository.

Personal Memory and Procedure Memory are global to the user under $MEMORAX_CODE_HOME/personal-memory/ (default ~/.memorax-code/personal-memory/): User Profile uses user-profile/preferences.md, and each Procedure topic uses its own file under procedure-memory/. Applicability may mention a repository, tool, or workflow, but no personal-memory layer exists inside a repository. Existing .repo_memory personal-memory files are ignored and are not migrated. The Agent saves by intent rather than wording: when you state a working rule or preference meant to keep applying beyond the current task, it saves it without waiting for "remember" and tells you what it stored; rules about how to work go to Procedure Memory, and communication or presentation preferences go to the User Profile. When saved content already exists, MemoraX Code compares its meaning before writing: an equivalent request makes no change; a durable refinement or conflict updates the matching entry and removes the superseded wording; an invalid scope is corrected, or the entry is deleted only when it is wholly obsolete. An explicit forget request deletes only the named preference, procedure topic, section, or step and leaves unrelated memory unchanged. One-time task instructions do not change saved memory, and the Agent asks before writing when the durable intent or target is unclear.

MemoraX is required for cloud-backed memory. Completing setup activates MemoraX search/add and the generated configuration's automatic writeback; there is no second writeback confirmation. Search runs when the agent uses the Skill or you invoke the CLI. Hooks provide local memory context and reminders without issuing Search requests.

Optional Jev configuration lets a separate TypeSafe model judge each eligible distinct user request. When Search is useful, the agent reads the memorax-code Skill's Search reference and follows its query and execution guidance. It sends bounded current-request and previous-turn text to TypeSafe, uses your Jev key, and is disabled by default. Failures fall back to the normal Skill reminder cadence; the agent still executes Search.

Local trace capture is enabled by default for supported clients. Depending on client capabilities, retained traces under MEMORAX_CODE_HOME may contain prompts, responses, recalled memory, reminder text, and local paths. Use the local trace settings to switch to metadata-only capture or disable a client's trace. Minimal local session state remains when trace is disabled so memory operations keep the correct workspace scope.

Failures in setup, explicit Search/Add, Backend start/stop/restart, client deployment, and package updates provide an error code, recovery guidance, and a local content-free diagnostic; see failure recovery. Hook execution and automatic writeback also retain known failures locally with Debug off, without adding messages to the conversation or blocking the coding task. memorax-code status summarizes recent failures; use memorax-code logs --diagnostics or memorax-code logs --id <diagnostic-id> for details you can review and share.

Coding Memory follows the repository or workspace. Recognized default chat directories or contexts in Codex, WorkBuddy, OpenCode, and Cursor share General under the same configured MemoraX user ID. A Cursor conversation with no folder selected uses this shared @General scope; opening a repository uses that repository's normal scope. Existing memories are not migrated; see memory scope for the directory rules.

Guest quota reminders may display the complete Mark ID. Treat reminder text and retained traces containing it as sensitive.

Active memory operations send their query or selected content to MemoraX. Automatic writeback sends selected user instructions and the matching final Agent response from trusted workspace turns after removing the final-answer memory-impact disclosure, then extracts and stores reusable memory. It does not upload the complete retained client trace artifact or local trace path.

QA writeback preserves available native timestamps and labels observation-time fallbacks; see message timestamps.

Sign in to MemoraX Console at any time to view, edit, or delete saved memories. MemoraX Cloud does not receive model-provider credentials or local Backend tokens.

Read Configuration for all settings and Security for network, local-data, and retention boundaries.

For a global npm installation:

memorax-code update

Setup also enables background updates while the managed Backend is running. An update briefly stops a running managed Backend and restores it with the retained client selection. Setup after reinstall and updates preserve explicit client choices even when older configuration omits codex or claude. See update settings for release channels, custom state roots, client selection, Backend restoration, and disabling background checks. Restart or refresh a client after an update changes integration assets it has already loaded. If package replacement fails, follow the update recovery steps for memorax-code update --recover.

For older Windows installations that fail with an EBUSY rename error, follow the legacy upgrade recovery.

Run the product lifecycle before removing the npm package:

memorax-code uninstall

This removes managed integrations and the global package while retaining configuration and stored memories. Do not run npm uninstall -g first: it can remove the product command before client integration cleanup runs. See Uninstall and Retention for the complete retained-data list.

After a complete uninstall and reinstall, run memorax-code setup again; default setup reuses a complete retained connection. A normal memorax-code stop or partial client uninstall preserves setup completion.

• Installation and first use
• Configuration
• Troubleshooting
• Contributing
• Security

Issues and pull requests are welcome. Read CONTRIBUTING.md before making a change, and never include API keys, raw transcripts, private memory, or local trace artifacts in a public report.

MemoraX Code is available under the MIT License.