Sediment

Semantic memory for AI agents. Local-first, MCP-native.

Combines vector search, a relationship graph, and access tracking into a unified memory intelligence layer — all running locally as a single binary.

Why Sediment?

• Single binary, zero config — no Docker, no Postgres, no Qdrant. Just sediment.
• 50ms store, 103ms recall — local embeddings and vector search at 1K items, no network round-trips.
• 4-tool focused API — store, recall, list, forget. That's it.
• Works everywhere — macOS (Intel + ARM), Linux x86_64. All data stays on your machine.

Comparison

Install

# Via crates.io
cargo install sediment-mcp

# Via Homebrew
brew tap rendro/tap
brew install sediment

# Via shell installer
curl -fsSL https://raw.githubusercontent.com/rendro/sediment/main/install.sh | sh

# From source
cargo install --path .

Setup

Add Sediment to your MCP client configuration:

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "sediment": {
      "command": "sediment"
    }
  }
}

Claude Code

Run sediment init in your project, or add manually to ~/.claude/settings.json:

{
  "mcpServers": {
    "sediment": {
      "command": "sediment"
    }
  }
}

Cursor

Add to .cursor/mcp.json in your project:

{
  "mcpServers": {
    "sediment": {
      "command": "sediment"
    }
  }
}

VS Code (Copilot)

Add to .vscode/mcp.json in your project:

{
  "servers": {
    "sediment": {
      "command": "sediment"
    }
  }
}

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "sediment": {
      "command": "sediment"
    }
  }
}

JetBrains IDEs

Go to Settings > Tools > AI Assistant > MCP Servers, click +, and add:

{
  "sediment": {
    "command": "sediment"
  }
}

Zed

Add to ~/.zed/settings.json on macOS or ~/.config/zed/settings.json on Linux:

{
  "context_servers": {
    "sediment": {
      "command": "sediment",
      "args": [],
    }
  }
}

Tools

CLI

sediment                          # Start MCP server
sediment init                     # Set up Claude Code integration
sediment stats                    # Show database statistics
sediment list                     # List stored items
sediment list --scope global      # List global items
sediment store "some content"     # Store content
sediment store -                  # Store content from stdin
sediment recall "search query"    # Search by semantic similarity
sediment forget <id>              # Delete an item by ID

All commands support --json for machine-readable output.

Environment Variables

Example — use a throwaway database for a task:

SEDIMENT_DB=/tmp/task-xyz sediment store "temporary context"
SEDIMENT_DB=/tmp/task-xyz sediment recall "context"

The --db CLI flag takes the same value and overrides the env var if both are set.

How It Works

Two-Database Hybrid

All local, embedded, zero config:

• LanceDB — Vector embeddings and semantic similarity search
• SQLite (access.db) — Relationship graph, access tracking, decay scoring, consolidation queue

Key Features

• Memory decay: Results re-ranked by freshness (30-day half-life) and access frequency. Old memories rank lower but are never auto-deleted.
• Trust-weighted scoring: Validated and well-connected memories score higher.
• Hybrid search: Vector similarity combined with FTS/BM25 scoring for better retrieval quality.
• Project scoping: Automatic context isolation between projects. Different-project items receive a similarity penalty.
• Relationship graph: Items linked via RELATED, SUPERSEDES, and CO_ACCESSED edges. Recall expands results with 1-hop graph neighbors and co-access suggestions.
• Background consolidation: Near-duplicates (≥0.95 similarity) auto-merged; similar items (0.85–0.95) linked.
• Type-aware chunking: Intelligent splitting for markdown, code, JSON, YAML, and plain text.
• Conflict detection: Items with ≥0.85 similarity flagged on store.
• Cross-project recall: Results from other projects flagged.
• Local embeddings: all-MiniLM-L6-v2 via Candle (384-dim vectors, no API keys).
• Model integrity: SHA-256 verification of all model files on every load, pinned to a specific revision.
• Auto-migration: Database schema automatically migrated when upgrading from older versions.

Security

• Input bounds: Content (1MB), queries (10KB), JSON-RPC lines (10MB).
• Rate limiting: 600 tool calls per minute.
• SQL injection prevention: Sanitized filter expressions for LanceDB; parameterized queries for SQLite.
• Cross-project access control: Forget enforces project isolation. Cross-project content is flagged in recall results.
• Error sanitization: Internal errors logged to stderr; only generic messages returned to MCP clients.
• Retry with backoff: Transient failures retried with exponential backoff (3 attempts, 100ms–2s).

Performance

Benchmarked against 5 alternatives with 1,000 memories and 200 queries. See BENCHMARKS.md for full results.

Data Location

• Vector store: ~/.sediment/data/
• Graph + access tracking: ~/.sediment/access.db

Everything runs locally. Your data never leaves your machine.

Contributing

See CONTRIBUTING.md for build instructions and PR guidelines.

License

MIT