open-source memory runtime · Apache-2.0 · v1.3.2
Memory your agents can show their work on.
dreamd is a local-first memory runtime for AI agents. What one agent learns, the next one already knows, and every recall can be traced to the memory that drove it.
The memory is plain files in your repository. You can read it, diff it, and remove a bad lesson with a receipt.
{
"schema_version": "1.0.0", "id": "evt_01JR04101000XYZB2C3D4E5F6G", "timestamp": "2026-04-10T10:00:00Z", "pain": 3.5, "importance": 4.5, "pinned": false, "skill_action": "rust::async::stream_buffered", "source_harness": "aider", "content": "futures::stream::iter(urls).map(fetch).buffered(8) caps concurrent in-flight tasks at 8 while preserving output order. Switch to buffer_unordered(8) when downstream is order-independent, same backpressure, lower tail latency."} 01 the memory is a file
One folder, shared by every agent you run.
Drop a .agent/ folder in the project. Claude Code, Cursor and Cline read and write it over MCP. The episodic log is JSONL, the consolidated lessons are Markdown, and both are meant to be committed with your code.
.agent/
├── episodic/
│ └── AGENT_LEARNINGS.jsonl # append-only observation log
├── semantic/
│ └── LESSONS.md # dream-cycle output
├── personal/ # your preferences and decisions
├── working/ # in-flight task context
└── .dreamd/ # daemon state (gitignored) Most memory tools treat a hosted database as the source of truth and files as an export. dreamd inverts that. You can cat, grep and git diff every byte, and the memory stays readable without dreamd installed.
02 glass box
When a memory is wrong, you can find it and take it back.
-
Trace
dreamd blame shows which stored memories drove a recall, with id, timestamp, score and layer for each.
dreamd blame <query> -
Remove
dreamd forget removes one event and the lesson state that names it, and can write a receipt of what it removed.
dreamd forget <event-id> --proof receipt.json -
Verify
A provenance ledger links lessons to the events they came from. doctor recomputes its Merkle root and lists broken edges.
dreamd doctor --provenance
03 where it stands
Shipped, versioned, and public.
- release
- v1.3.2 18 November 2026
- license
- Apache-2.0 open source, self-hosted, no telemetry
- mcp tools
- 2 search_nodes and append_node
- adapters
- 3 Claude Code, Cursor and Cline
- runtime
- Rust single binary, about 20 MB stripped
- platforms
- Linux · macOS · Windows Windows over loopback TCP
04 direction
From audit to trust.
What ships today is agent memory you can read, review and roll back. Every lesson cites the events it came from, each dream cycle lands as a commit you can review in a pull request, and a bad memory comes out with dreamd forget or a git revert.
The roadmap goes from auditing memory to trusting it: trust labels, so a lesson from an untrusted source can't drive a privileged action; signed forget receipts; and published results on whether learned lessons actually improve task success over time.
Robots and edge devices come later. dreamd is one local binary with no cloud dependency, the shape an offline robot needs, but no robot integration ships today. Roadmap items are direction, not dated commitments.
05 install
Two commands in a project root.
npx -y dreamd-mcp setup # scaffold .agent/ and wire your harness
npx -y dreamd-mcp # the MCP server your harness spawns
Requires Node.js 20 or newer. Prebuilt binaries for Linux x86_64 and macOS; on Windows, use a cargo build or a release asset.
06 team
Who builds dreamd.
- Phillip Austin Green
Founder. Builds the dreamd runtime.
- Joshua Beeney
Cofounder and CGO. Marketing and design.
dreamd is a botzr Corp company. It is not affiliated with Dream Labs.
Investors and partners: hello@dreamd.dev