Agent context that improves between sessions.

Open, local-first memory for AI agents.

Capture what happened, retrieve what matters, and review what changed. OpenDream helps agents carry useful context across projects, tools, and sessions while keeping boundaries, sources, and review visible.

Closed products are starting to dream. OpenDream makes agent memory open, local, portable, and reviewable. "Dreaming" means background memory review and cleanup, not an opaque model behavior.

The observability UI shows recent agent activity, selected memories, context-use audit records, review decisions, and workspace health without making you read raw JSON first. Full light and dark demo cuts: docs/showcase/ui-demos.md.

Multiple agents, one workspace memory plane

Codex, Claude Code, Cursor, and other agents can contribute across sessions in the same workspace. OpenDream keeps those events and prepared contexts in one local store with agent attribution, so a later agent can inspect and reuse what came before without replacing its built-in memory.

Watch the full session and diff flow: light MP4, dark MP4, or the UI demo gallery.

OpenDream memory in real agent sessions

Current status: OpenDream is alpha software. Codex is currently the most tested path; other integrations depend on each host's hooks, rules, context files, or CLI surface. Current evals are fixture-driven release checks, not broad real-world benchmarks. Memory can hurt when it is stale, over-broad, or incorrectly scoped, so reviewability is part of the product goal rather than a finished claim.

Quick start

uv tool install opendream   # or: pipx install opendream
opendream init --workspace .
opendream verify activation-capture --workspace . --targets configured
opendream status --workspace .
opendream deactivate --workspace .

Run the self-contained showcase in a disposable workspace:

opendream eval showcase --scenario agent-workspace-showcase --workspace .tmp/opendream-showcase
opendream observe serve --workspace .tmp/opendream-showcase

If your installed CLI does not show activate, verify, deactivate, semantic, or eval, upgrade with uv tool upgrade opendream or reinstall from Git. After upgrading, opendream verify --help, opendream semantic --help, and opendream eval --help are quick checks that your install matches the docs.

Bleeding-edge from Git (overwrites the tool env): uv tool install --force "opendream @ git+https://github.com/pylit-ai/opendream.git".

Install checks, upgrade notes, known limitations, and clean-room provenance live in docs/launch-readiness.md, docs/known-limitations.md, and docs/provenance.md.

CLI demos for docs and operators

OpenDream's first-run path is intentionally small. These clips are useful beside command reference docs; the UI application tour above is the better first look.

Demo	What it shows
	init with default activation, status, and deactivation in a local workspace.
	prepare-context turns durable repo memory into prompt-ready agent context.
	Unrelated prompts return an explicit no-match instead of injecting stale or irrelevant memory.
	The showcase eval compares stateless vs memory-assisted answers and keeps negative controls in the report.

UI demo gallery: docs/showcase/ui-demos.md. CLI demo gallery: docs/showcase/cli-demos.md. Brand motion assets: frontend/assets/animation/README.md.

uv tool install opendream
# or: pipx install opendream
opendream --help

uv tool install "opendream @ git+https://github.com/pylit-ai/opendream.git"
# or: pipx install git+https://github.com/pylit-ai/opendream.git

make setup
.venv/bin/opendream --help

# macOS
brew install jq

# Ubuntu / Debian
sudo apt-get install jq

# Other systems
# See: https://jqlang.github.io/jq/download/

Integration at a glance

OpenDream is an activation-first CLI.

Codex is currently the most tested path. Other integrations expose context through CLI commands, rules, hooks, or generated files depending on host support. A built-in target means OpenDream can generate the integration surface; it does not mean every host executes hooks natively.

You do not need to replace existing agent memory. OpenDream keeps a separate workspace-local store and can run alongside built-in memory from Codex, Claude Code, or another agent. Activation adds OpenDream-managed integration surfaces; preview them with activation-plan, skip them with init --no-activate-configured, or remove them later with deactivate. See the FAQ.

opendream init --workspace .
opendream verify activation-capture --workspace . --targets configured
opendream status --workspace .
opendream activate --workspace . --repair
opendream deactivate --workspace .

The lower-level runtime remains available, but it is not the main mental model.

Agent integration details (workspace vs cwd, memory_layout, empty_reason / hints, JSON version): docs/agent-integrations.md.

Recommended activation workflow

1. opendream init --workspace . — create the memory layout and activate configured agent surfaces by default.
2. opendream activation-plan --workspace . --targets configured — optional dry-run: see which surfaces are detected. Use --targets all-supported to preview every built-in agent target.
3. opendream verify activation-capture --workspace . --targets configured — run generated hooks in the selected workspace and prove diagnostic memory capture.
4. opendream status --workspace . — report the latest capture verification state (never_run, passed, or failed) with the next action.
5. opendream activate --workspace . --repair — restore drifted managed files and hook entries.
6. opendream doctor --workspace . --surface agents — inspect managed surfaces without writing diagnostic memory.

Compatibility matrix

For a tool that was not detected yet, run e.g. opendream activate --workspace . --targets cursor once to create .cursor/rules/opendream.mdc and hook scripts.

Instruction-only targets (Cursor rules, .github/copilot-instructions.md, .hermes.md) ship the same generic pre/post shell hooks as the Codex example target; the agent must still run those commands when the host has no native OpenDream hooks.

Corrections worth knowing:

• Treat maintain as the documented maintenance entrypoint even if the CLI exposes more commands.
• First-party surface = this CLI. Hook/script glue is operator-owned unless you add it.
• No first-party MCP server in this repo; docs/mcp/servers.md is a template for inventorying MCP, not a shipped server.

Workspace, Runtime, and Observability

OpenDream stores canonical state per workspace under .opendream/. The multi-workspace catalog is a derived, machine-local convenience index; workspace memory stays the source of truth and scans only run on roots you configure.

opendream workspace list
opendream observe serve --workspace "$PWD" --port 8000

Semantic-first mode is a configuration posture, not an automatic readiness claim. If the selected semantic path cannot run, OpenDream reports a degraded state with a reason and next action; prepare-context should still use progressive disclosure and pruning evidence so recalled context stays inspectable.

Use the detailed operator guide for workspace catalog setup, upgrade/backfill flows, runtime commands, generated data locations, global/project stores, semantic mode, automation, eval notes, and full CLI examples: docs/operator-guide.md.

Documentation

OpenDream does not send telemetry by default. Provider/API-key paths are operator-configured execution paths and are separate from analytics or tracking.

Contributing

For local development, run:

make sync
make verify
make release-check

Contributor workflow, smoke tests, automation examples, verification targets, and maintainer release notes live in docs/contributor-reference.md. Security reports go through SECURITY.md; general contribution guidance lives in CONTRIBUTING.md.