# Universal Memory (UMem) [](https://pypi.org/project/universal-memory/) [](https://pypi.org/project/universal-memory/) [](https://github.com/YanAmorelli/universal-memory/blob/dev/LICENSE) **[Website](https://universal-memory.com)** | **[Documentation](https://docs.universal-memory.com)** A vendor-agnostic cognitive persistence layer for AI agents. Eliminate the "repetition tax" by transporting your context, preferences, guidelines, and history seamlessly across sessions, IDEs, and LLM models. To see the core idea visually, check out the [Excalidraw design](https://excalidraw.com/#json=j3XjQIWMYEnkIzHpypuBb,rNJaVOECDGZ3WSuEYcCDjQ) or the proposal structure:  ### Diagram Breakdown * **Short-Term Memory (Ephemeral):** Project-specific (folder-level) memories. A simple summary of recent changes, pending tasks, and project or task-level constraints. * **Agents Behaviours:** Comports the user's expected agent behaviors. Instead of requesting the same settings in every session, the agent understands the user by their traits, thoughts, and any context key to enhancing the overall experience. This encompasses: * **Long-Term Memory** * **Short-Term Memory** * **User Preferences** * **Skill Creator:** Encapsulates understanding of specific workflows. When a user explains a task pattern multiple times, the system translates it into structured, reusable agent skills. * **Unified Instruction File (`AGENTS.md`):** The shared persistence endpoint consumed by compatible local agent instances (e.g., Agent A, Agent B, Agent C). --- ## The Problem: The "Repetition Tax" Every time you open a new session in Claude Code, start a new chat in Cursor, spin up a terminal with OpenCode, or invoke a local AI assistant, you pay a steep cognitive tax: * Re-explaining your stack (e.g., "We use Python 3.12, Typer, and Ruff"). * Repeating coding style preferences (e.g., "Prefer functional design, do not write docstrings unless requested"). * Copy-pasting database connection schemas or module layouts. * Explaining workflow methodologies (e.g., "We follow Spec-Driven Development (SDD)"). Universal Memory acts as a local persistence layer that automatically connects to your AI runtimes, aligning them to your exact workflow, context, and rules with zero friction. --- ## Key Architectural Concepts ### 1. Dual-Memory Model * **Short-Term Memory (Project Scope):** Ephemeral, directory-specific context. Tracks what you did 10 minutes ago, current active tasks, and immediate constraints. * **Universal Memory (Global Scope):** Long-lived preferences, style guidelines, tool configurations, and identity. ### 2. Auto-Adaptation Engine Instead of copy-pasting instructions, `umem` monitors your session context and automatically updates active project instruction manifests (`AGENTS.md`, `CLAUDE.md`, `.cursor/rules/`, etc.), enforcing operational consistency across all agents. ### 3. Model Context Protocol (MCP) Integration Integrate `umem` natively with any client supporting the standard MCP (such as Claude Desktop or Cursor). AI agents can programmatically retrieve context, learn new facts, and suggest skills on the fly. ### 4. Agent Skills Standard Encapsulates complex, repetitive procedural instructions into formal Agent Skills (conforming to the [agentskills.io](https://agentskills.io) standard), complete with structured directories containing `SKILL.md` instructions, helper `scripts/`, and documentation `references/`. Universal Memory keeps one canonical source for each skill. Shared, user-facing project skills live under `umem/skills//SKILL.md`; private, operational, and legacy project skills live under `.umem/skills//SKILL.md`. Native runtime folders such as `.agents/skills/`, `.opencode/skills/`, and `.antigravity/rules/` receive complete synchronized copies so each agent can consume the same skill in its expected layout. --- ## Installation & Setup Ensure you have Python 3.12+ installed. You can run or install `umem` using your preferred package manager. ### Try instantly with `uvx` You can run `umem` without installing it permanently: ```bash uvx --from universal-memory umem --help ``` > [!WARNING] > `uvx` is best for quick trials. For ongoing use, install Universal Memory as a persistent tool so `umem` is always available and can fully manage long-lived global memories and synced agent skills: > ```bash > uv tool install universal-memory > ``` ### Install via PyPI ```bash pip install universal-memory ``` ### Upgrade Universal Memory `umem update` does not upgrade the Python package from PyPI. It performs local, offline maintenance for the current `.umem` workspace, such as schema migrations, benchmark refreshes, and skill synchronization. To upgrade the installed `umem` executable, use the package manager that installed it: ```bash # If installed with uv tool uv tool upgrade universal-memory # If installed with pipx pipx upgrade universal-memory # If installed with pip python -m pip install --upgrade universal-memory # If running temporarily with uvx uvx --refresh --from universal-memory umem --version ``` Confirm the executable you are running: ```bash umem --version which umem ``` Upgrading the executable does not silently mutate existing projects. The next time you work in an initialized project, reconcile it locally: ```bash umem update --check umem update umem update --skills umem connect umem doctor ``` You do not need to run `umem init` again. Local maintenance creates snapshots and audit records before UMEM-owned writes. Existing `.umem/skills/use-universal-memory/` trees and customized managed files are preserved; if both legacy and canonical Universal Memory skill roots exist, UMEM stops for an explicit migration decision instead of merging or deleting either tree. --- ## Quick Start Guide ### 1. Initialize your project Open your project directory and run: ```bash umem init ``` Universal Memory detects the agents already used in the workspace, presents one combined confirmation, configures the best available project integration, and verifies that the agent can read project context. You do not need to choose an integration mechanism or know which instruction files it uses. When a compatible agent needs the portable Agent Skill, UMEM discloses any network use and external project-scoped copy before confirmation, disables anonymous installer telemetry, and treats a missing prerequisite or failed installation as recoverable instead of blocking initialization. To connect another agent later, run: ```bash umem connect ``` Explicit runtime selection remains available for automation and unusual setups, but it is not required for the normal path.

How the portable Tier 2 installation works

UMEM resolves the detected agent's project skill directory from a reviewed catalog pinned to `skills@1.5.20`, runs one project-scoped installation, and validates the complete installed skill tree plus a real `umem context` read. It does not install into a second project and copy the result back. The command orchestrated by UMEM in `v0.6.2` is equivalent to: ```bash DISABLE_TELEMETRY=1 npx --yes skills@1.5.20 add https://github.com/YanAmorelli/universal-memory/tree/v0.6.2/skills/universal-memory --skill universal-memory --agent pi --copy -y ``` Here `pi` is an example; UMEM supplies the detected agent ID. Node.js and `npx` are optional prerequisites for this external bridge. When either is unavailable, initialization remains usable and UMEM reports a managed or manual fallback. Unknown agent IDs never execute `npx`.

### 2. Bootstrap one agent session At the start of each conversation or agent session, prefer the MCP `bootstrap()` tool when it is connected. Otherwise, use the equivalent CLI command: ```bash umem bootstrap --format json ``` This single call validates the integration and returns project status, active project context, and the skills catalog. Treat `data.context` as active context, inspect `data.skills.list`, and request details only for skills relevant to the current task: ```bash umem skills detail --format json ``` Run bootstrap only once per conversation or session. It replaces the previous startup sequence of separate `status`, `context`, and `skills list` calls; it does not perform installation, synchronization, or configuration. ### 3. Save your first preferences and facts Tell `umem` what to keep in mind. You can target either the project scope (this folder) or the global scope (across all projects): ```bash # Save a global preference umem remember --scope global "Yan is a solutions architect specializing in AI applications" # Save a project-specific constraint umem remember --scope project "Always use Tomllib instead of PyYAML for configuration files" --tag config ``` ### 4. Retrieve Context Verify the consolidated context summary generated by combining short-term facts, rules, and global preferences: ```bash umem context --scope project ``` ### 5. Adopt or create an Agent Skill If a skill already exists, choose the safest adoption path first. Use `adopt` for an existing `.umem/skills/` directory; use `import` for native runtime directories such as `.agents/skills/` and sync it back out to configured runtimes: ```bash umem skills adopt .umem/skills/review-protocol --scope project umem skills import .agents/skills/review-protocol --scope project --sync umem skills detail review-protocol ``` If you are starting from scratch, draft and publish it without native side effects: ```bash umem skills draft create \ --name "Review Protocol" \ --description "Reusable review workflow" \ --trigger "when reviewing code" umem skills draft validate review-protocol umem skills publish review-protocol --format summary ``` For a one-step workflow, create the canonical skill. It is canonical-only by default; request sync explicitly when native runtime targets should be written: ```bash umem skills create \ --name "Review Protocol" \ --description "Reusable review workflow" \ --trigger "when reviewing code" \ --format summary umem skills sync review-protocol --check-gitignore --format summary ``` After editing `.umem/skills/review-protocol/SKILL.md`, refresh one runtime skill with: ```bash umem skills sync review-protocol ``` ### 6. Check status and health ```bash umem status ``` --- ## Host Integration & Support Matrix UMEM deliberately separates native ownership from portable compatibility: | Tier | Contract | Guarantee | | --- | --- | --- | | **Tier 1 — Native/Managed** | Maintained host adapter, native setup and repeatable validation | UMEM owns and tests the documented integration. | | **Tier 2 — Directed CLI** | `AGENTS.md` or the official Agent Skill directs a shell-capable agent to the UMEM CLI | UMEM validates portable instructions, CLI access, and context reading, but not every host-specific behavior. | | **Tier 3 — Unmanaged MCP** | The user manually connects MCP to a host without a programmed workflow | UMEM validates MCP availability only; agent behavior is not guaranteed. | The maintained and named integration surfaces are: | Runtime / Host | Support Tier | Config / Instructions Target | | --- | --- | --- | | **Claude Code** | Tier 1 — Native/Managed | `CLAUDE.md`, `.claude/skills/`, `.claude/settings.json` | | **OpenCode** | Tier 1 — Native/Managed | `AGENTS.md`, `.opencode/skills/`, `.opencode/opencode.jsonc` | | **Codex (OpenAI)** | Tier 1 — Native/Managed | `AGENTS.md`, `.agents/skills/`, `.codex/config.toml` | | **Cursor** | Tier 2 — Directed CLI | `.cursor/rules/universal-memory.mdc` | | **Antigravity** | Tier 2 — Directed CLI | `.antigravity/rules/universal-memory.md` | | **Pi, Gemini CLI, GitHub Copilot, Cline, Zed, and other reviewed Agent Skills hosts** | Tier 2 — Directed CLI | Project skill directory pinned to the `skills@1.5.20` catalog | | **Windsurf** | Tier 2 — Frozen legacy adapter | `.windsurf/skills/universal-memory/` | | **Unmodeled MCP host** | Tier 3 — Unmanaged MCP | User-managed MCP configuration | An agent appearing in the external `skills` catalog does not make it Tier 1. Tier 1 is intentionally small and requires a maintained adapter, release evidence, and repeatable host-specific validation. See the [Getting Started guide](docs/users/getting-started.md) for legacy-project behavior and the portable installation flow. --- ## Running as a Model Context Protocol (MCP) Server AI agents can interact directly with your memory over the Model Context Protocol. Manual MCP configuration for a host without a programmed UMEM workflow is Tier 3: tool availability is validated, but instruction loading and agent behavior are not guaranteed. ### One-off Launch Command ```bash uvx --from universal-memory umem-mcp ``` ### Persistent Install Launch Command ```bash umem-mcp ``` ### Bootstrap Once Per Session When the MCP server is connected, agents should call `bootstrap()` once at the beginning of the conversation or session. It is semantically equivalent to `umem bootstrap --format json`: both return status, active project context, and the skills catalog, and both preserve the same fail-fast error behavior. Skill details remain separate and should be requested only for selected relevant skills. ### Example Config: Claude Desktop (`claude_desktop_config.json`) Use the `uvx` form when Universal Memory is not installed as a persistent tool: ```json { "mcpServers": { "universal-memory": { "command": "uvx", "args": [ "--from", "universal-memory", "umem-mcp" ] } } } ``` If you installed Universal Memory with `uv tool install universal-memory` or `pipx install universal-memory`, use the stable entrypoint: ```json { "mcpServers": { "universal-memory": { "command": "umem-mcp", "args": [] } } } ``` Troubleshoot startup with: ```bash uvx --from universal-memory umem doctor uvx --from universal-memory umem-mcp --help ``` For GUI-launched MCP hosts, use the absolute path to `uvx` if the host does not inherit your shell `PATH`. --- ## Safety & Guardrails * **API Secret Scanner:** `umem` passes all incoming facts through a passive scanner to block API keys, tokens, or credentials from being stored in your persistent cognitive base. * **Snapshots & Rollbacks:** Every automated update to your config files (`AGENTS.md`, `CLAUDE.md`) is preceded by a snapshot backup. You can rollback anytime: ```bash # View audit logs umem audit list --scope project # Revert last automated modification umem rollback --scope project ``` * **Skill Drift Protection:** `umem skills sync` detects managed native drift and keeps local changes by default. Use `--drift-decision overwrite` only when you intentionally want canonical UMEM content to replace the managed native copy. * **External Bridge Boundary:** Tier 2 installation through `npx skills` is an explicitly confirmed external mutation. UMEM disables anonymous installer telemetry, constrains the target to the current project, and validates the complete result, but labels the write as externally executed rather than claiming UMEM snapshot ownership. --- ## Managing Agent Skills You can draft, create, adopt, import, validate, maintain, and sync specialized behaviors: ```bash # List all active skills umem skills list # Inspect one skill umem skills detail review-protocol # Draft, validate, and publish without native runtime writes umem skills draft create --name "Review Protocol" --description "Reusable review workflow" umem skills draft validate review-protocol umem skills publish review-protocol # Create a new canonical skill and explicitly sync native targets umem skills create --name "Review Protocol" --description "Reusable review workflow" --sync # Adopt existing canonical work umem skills adopt .umem/skills/review-protocol --scope project # Import an existing native skill and distribute complete runtime copies umem skills import .agents/skills/review-protocol --scope project --sync # Validate and maintain canonical skills umem skills validate review-protocol umem skills canonical update review-protocol --file .umem/skills/review-protocol/SKILL.md umem skills rename review-protocol --slug review-checklist umem skills cleanup review-checklist --targets --format summary umem skills cleanup review-checklist --targets --apply umem skills repair --remove-orphan-targets --format summary # Synchronize one canonical skill into active native runtime folders umem skills sync review-protocol --check-gitignore --format summary # Synchronize all active canonical skills during maintenance umem update --skills # Track and review recurring workflow candidates umem skills track --name "Review Protocol" --description "Recurring review workflow" umem skills recommend --scope project umem skills propose --decision yes umem skills promote --yes umem skills generate --yes ``` --- ## License Distributed under the Apache License 2.0. See [LICENSE](LICENSE) and [NOTICE](NOTICE) for more information.