dreamd compared
Intent. These are architecture notes, not a scoreboard. They describe how the systems differ in shape: where data lives, what the setup looks like, what the data model is. Which system actually holds up when facts change is an empirical question, and it is the one the State-Drift Benchmark answers. Results publish October 9, 2026; every system on this page runs the same scenarios, dreamd included.
vs manual CLAUDE.md / AGENTS.md
Writing memory by hand works until you have more than one agent, more than one session, or more than one repository.
| Manual markdown files | dreamd | |
|---|---|---|
| Updates across tools | You copy-paste | Automatic, same folder |
| Structured memory | Whatever you write | Frontmatter schema + types |
| Consolidation | You edit by hand | Salience-promoted from logs |
| Works offline | Yes | Yes |
| Harness-agnostic | Only if you maintain N files | Single .agent/ folder |
dreamd doesn’t replace what you wrote down; it captures what you didn’t. The AGENTS.md you maintain by hand is the anchor. dreamd fills in the implicit learning your agent accumulates between sessions.
vs Mem0
Mem0 is a hosted memory layer. dreamd is a local daemon. Same job, different shape.
| Mem0 | dreamd | |
|---|---|---|
| Architecture | SaaS API, cloud-hosted | Single binary, local daemon |
| Data location | Their servers | Your .agent/ folder |
| Recall path | Network round-trip | Local filesystem |
| Offline capable | No | Yes |
| Open source | Partial (enterprise features closed) | Full (Apache-2.0) |
| Harness support | OpenAI SDK shape | Claude Code, Cursor, Cline native |
The difference is where your data lives and who controls the pipeline, not which one remembers better. That question goes to the benchmark, where both systems run the same scenarios.
vs Letta (formerly MemGPT)
Letta is an agent operating system with managed memory. dreamd is a memory daemon for existing agents.
| Letta | dreamd | |
|---|---|---|
| What it is | Agent OS + managed memory | Memory layer for your existing tools |
| Your agent | Letta’s built-in agent | Claude Code, Cursor, Cline, whatever you already use |
| Data model | Letta’s internal state | .agent/ folder in your repo |
| Self-hostable | Yes (MemGPT was) | Yes, and it’s the only way it runs |
| Setup | New tool to learn | npx dreamd-mcp, point to repo |
If you want a new agent platform, Letta is built for that. If you want the agents you already use to share memory, dreamd fits that shape. Neither framing says anything about memory quality; that’s measured, not asserted.
vs building your own vector DB pipeline
You could wire up LanceDB, Chroma, or SQLite yourself. Many teams do.
| DIY vector pipeline | dreamd | |
|---|---|---|
| Salience formula | You design and tune | Built-in, configurable |
| Folder structure | You define | Specified, harness-agnostic |
| MCP server | You write | Shipped |
| Maintenance | You own | Maintained upstream |
If your memory needs are deeply custom, DIY makes sense. For the common case (save context, recall lessons, share across tools) dreamd ships the convention and the daemon together so you don’t have to invent both.
What this page deliberately doesn’t say
No latency numbers, no accuracy claims, no “X is better.” Those claims belong in the State-Drift Benchmark results, where every system, dreamd included, is scored by the same open-source harness on the same scenarios. Until then, this page sticks to architecture.
Comparison data sourced from published docs and READMEs as of v0.1. Claims are verifiable; if something is wrong, open an issue.