Memory in the workspace
Activate OpenDream inside the workspace you already use. Canonical state stays local, under the workspace folder, without forcing a hosted memory service into the loop.
Open, local-first memory for AI agents
Agent context that improves between sessions.
Capture what happened, retrieve what matters, review what changed. Carry useful context across projects, tools, and sessions while keeping boundaries, sources, and review visible.
Install
uv tool install opendream
Agent brief
Built for portable agent memory.
Use memory across agents and tools without hiding where it came from or where it applies.
Use when
cross-session repo memory
multi-agent repos
reviewable state
Supported
Claude Code
Codex
Gemini CLI
Copilot-style repo instructions
MCP workflows
Install
uv tool install opendream
opendream init --workspace .
Verify
opendream verify activation-capture --workspace . --targets
configured
opendream status --workspace .
Avoid
stateless RAG
forbidden durable state
unreviewed writes
Human Interface
See memory behavior before it becomes hidden context.
The browser UI is a readable control room for local memory: runs, retrievals, source links, evals, reviews, and workspace catalog all in one place.
Readable observability
Inspect recent runs, selected memories, prompt context, review decisions, and workspace health without reading raw JSON first.
Sleep-time maintenance
Dream runs turn agent and operator history into durable, source-linked context so future sessions can recover what mattered.
Skeptical by design
Source links, abstention paths, evaluation reports, and review queues make memory quality inspectable before you trust it.
Memory surface
Health, contested state, type mix, and recent durable memory highlights.
Retrieval log
Prompt-context assembly with query history and selected source links.
Prompt context
The exact context package a later session can inspect before model handoff.
Shared agent sessions
Codex, Claude Code, and Cursor can contribute across sessions while event trails remain attributable.
Run diff
Memory changes are shown as inspectable diffs before you treat them as truth.
Review queue
Recommended memory decisions before questionable context becomes trusted.
Relationship graph
Use the graph as a quick map of records, runs, retrievals, and context-use links.
Context evidence
Know what shaped the next prompt.
OpenDream shows which memories were selected, why they matched, and which items stayed out of the handoff.
Selected memories stay linked to their source
rows.
Context use is visible before it becomes agent
behavior.
Safety checks flag stale, low-confidence, or
contaminated memory.
Context improvements
Agent threads where context changed the answer.
Real coding-agent output, normalized for readability.
OpenDream recovered the previous state and pointed to the next concrete recovery step.
$ codex "recover current task state"
- Read memory index and prior run summary
› assistant
OpenDream found prior memory:
Polymarket merge done. Next: confirm clean
state, identify stash contents, create an
AAF15 recovery branch.
Remembered constraints changed the implementation path toward a read-only, rebuildable projection.
/resume implementation plan
* Searching repo constraints and saved context
› assistant
Next implementation I'd recommend:
Build a candidate index script that scans
artifacts and writes a Markdown summary. No
schema migration authority needed.
OpenDream context used: context_9d90bec8b454
Prior context triggered a verification step before editing, avoiding duplicated stale work.
$ codex "continue from accepted artifact"
OpenDream found related memory before any
edits.
› assistant
OpenDream surfaced prior project memory.
A similar milestone run may already have
completed. Verify current repo truth before
editing.
Demo videos
See how dreaming improves context.
Workspace overview
Current workspace overview
Workflow
From activity to reviewed context.
Capture
Agent activity lands as local events, transcripts, and source-linked project notes.
Dream
Background memory review and cleanup turns noisy history into saved, inspectable memory.
Retrieve
Task context pulls only memories that fit the current prompt or handoff.
Verify
Reports show selected source links, exclusions, changes, and safety checks.
Why OpenDream
Memory you can move, scope, and review.
| Capability | Static files | RAG / vector DB | Hosted APIs | Managed memory | OpenDream |
|---|---|---|---|---|---|
| Improves context across sessions | Manual | Depends | |||
| Local-first by default | Depends | ||||
| Portable across agent runtimes | Manual copy | Custom | API-dependent | Product-bound | |
| User-defined boundaries | Folder rules | Metadata | API-dependent | Product-defined | |
| Explains recalled context | Sometimes | Sometimes | Limited | ||
| Shows stale or conflicting memory | Limited | Limited | Limited | ||
| Reviewable memory updates | Manual diffs | Custom | Limited | Limited | |
| Works without vendor memory lock-in | |||||
| Public real-world benchmarks | Varies | Varies | Private | Not yet |
Full support
Partial or conditional
Manual setup
Not supported
In progress
OpenDream is alpha. Current evidence is strongest for local operation, context preparation, auditability, and Codex-tested workflows. Broader integrations and external benchmarks remain active validation areas.
Frequently asked
Add OpenDream without starting over.
Keep the agent setup and memory systems you already use. OpenDream adds a separate, inspectable memory layer.
Do I need to replace my existing memory system?
No. OpenDream keeps its own workspace memory and can run alongside built-in memory from Codex, Claude Code, or another agent. Use both, compare what each recalls, and decide where OpenDream adds value.
Will it overwrite my agent configuration?
Activation adds OpenDream-managed hooks, wrappers,
or marked instruction blocks for selected agents. It
does not replace their built-in memory. Preview
changes with activation-plan, or remove
generated integration with deactivate.
Can I try it without activating agent hooks?
Yes. Run opendream init --workspace .
--no-activate-configured, then use the CLI and
local UI directly before connecting an agent.
Does OpenDream send my workspace to a cloud service?
No. OpenDream does not send your workspace to OpenDream Cloud, and the current runtime has no OpenDream Cloud upload path. Future cloud features would be separately configured and opt-in. See the full FAQ for the separate status of model-provider paths.
OpenDream Cloud
Be the first to know about early access to OpenDream Cloud.
The local version stays 100% free and open source. Cloud access will open in small preview batches for hosted Pro, Team, and Enterprise workflows.