Skip to content

botzrDev/dreamd

dreamd

License: Apache 2.0 MCP-compatible Platforms Status

Same memory in every IDE.

dreamd makes Claude Code, Cursor, and Cline remember the same things. Drop a .agent/ folder in your repo. Every coding agent you use reads and writes to it.

Every coding agent ships its own memory format. dreamd is what they could share.

AGENTS.md is what you wrote down. dreamd is what your agent learned, across every tool.

The moment it earns its name

~/project $ npx dreamd-mcp init

# In Claude Code, Tuesday afternoon:
you   ▸ axum keeps blowing up when I unwrap in route handlers
claude▸ filed under rust::error_handling::axum_rejection

# In Cursor, Friday morning, fresh session:
you   ▸ why is this build failing?
cursor▸ You're unwrapping in a route handler. dreamd has a
        lesson from Tuesday -- axum needs IntoResponse on
        custom Error types. Try `?` and a typed error.

No re-explaining. No re-pasting. No "as I mentioned before."

What dreamd is -- and isn't

dreamd is dreamd isn't
A portable memory format (.agent/) checked into your repo A vector database
A reference MCP server for reading and writing it A knowledge graph engine
Local-first by default -- zero network calls without --llm A hosted SaaS
One source of truth across every coding agent you use A replacement for AGENTS.md or SKILL.md

If you need graph multi-hop reasoning, use Cognee. If you need a single-file portable memory capsule, use Memvid. dreamd does the one thing they don't: makes your memory follow you between coding agents.

Status

v0.1 in active development — targeting 2026-07-07. Sprint 3 in progress; 27 tickets shipped to date. The daemon builds and runs locally today: dreamd init, dreamd dream, dreamd doctor, dreamd mcp, dreamd watch, dreamd reset workspace, and dreamd version, plus the HTTP API (POST /api/v1/learn, GET /api/v1/recall, POST /api/v1/dream, GET /api/v1/preferences) on a Unix domain socket. The npx dreamd-mcp install path lands ahead of v0.1. Linux and macOS. See SPEC.md for the conformance contract and CONTRIBUTING.md to propose changes.

Star and Watch this repo to be notified when v0.1 lands.

Performance

Recall latency (warm in-RAM index, Criterion 0.5, WSL2/Linux):

Corpus size Mean (warm)
1 000 entries ~50 µs
10 000 entries ~313 µs
100 000 entries ~2.8 ms

Criterion reports mean across 100 samples; used here as the P50 proxy. All three sizes are well under the <5ms P50 warm NFR. Run cargo bench -p dreamd-core to reproduce.

Read-after-write visibility: up to commit_cadence_seconds (default 5s). A just-written event becomes recallable within one commit cycle; recall latency itself is unaffected.

Users who need tighter freshness can lower the commit cadence at the cost of higher index churn. User-facing config lands in v0.1.1; the cadence is a constructor argument today.

Getting started

Try it today (from source)

While v0.1 finishes baking, build the daemon locally:

git clone https://github.com/botzrDev/dreamd.git
cd dreamd
cargo install --path crates/dreamd-cli

# Then in any project:
cd ~/your-project
dreamd init      # scaffold .agent/
dreamd mcp       # speak MCP over stdio -- point Claude Code, Cursor, etc. at this

Requires Rust stable. See CONTRIBUTING.md for the full dev setup.

v0.1 install path

Landing ahead of 2026-07-07.

# scaffold .agent/ into the current project
npx dreamd-mcp init

# point Claude Code, Cursor, or any MCP-aware harness at the server
npx dreamd-mcp

Distribution: npm (primary). Cargo and Homebrew paths arrive in v0.1.1. See the v0.1 milestone.

Privacy

When LLM mode is enabled in v0.1.1, episodic content meeting the salience threshold may be sent to the configured provider. The personal/ layer is excluded unless you explicitly opt in. v0.1 makes no network calls. See SECURITY.md for the threat model and disclosure policy.

Platforms

v0.1 supports Linux and macOS. Windows lifecycle support arrives in v0.1.1.

Spec

The on-disk layout, JSON schema, salience formula, and dream-cycle contract are defined in SPEC.md. The spec is implementation-agnostic — dreamd is one implementation, not the only one.

To propose a change to the spec, open an issue prefixed with [RFC]. See CONTRIBUTING.md.

v0.1 progress

Layer Status
SPEC.md v0.1-draft Drafted
Reference implementation (dreamd daemon, HTTP API, dream cycle, Tantivy recall) In progress — Sprint 2 of 6
MCP server (dreamd mcp subcommand + npx dreamd-mcp shim) Subcommand built; npx distribution lands ahead of v0.1
CI / cross-platform matrix Lint, test, cross-platform build, binary-size gate, DCO check
Conformance test suite Not started

Contributing

See CONTRIBUTING.md for the development setup, commit conventions (DR-XXX story IDs), DCO sign-off policy, and the RFC process.

By participating in this project you agree to abide by the Code of Conduct.

To report a security issue, see SECURITY.md. Do not open a public issue for vulnerabilities.

License

Apache-2.0. See LICENSE and NOTICE.

About

Portable memory layer for AI agents -*pre-release

Topics

Resources

License

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages