anchored

README

Anchored

Persistent cross-tool memory for AI coding agents. Single binary. Zero dependencies.

[]

Anchored is an MCP memory server that gives Claude Code, Cursor, OpenCode, and any MCP-compatible tool a shared, persistent memory. Install once, and all your tools read, write, and search the same knowledge base.

No API keys. No daemon. All embeddings run locally.

Features

• Multilingual embeddings — paraphrase-multilingual-MiniLM-L12-v2 (50+ languages, PT-BR and EN parity)
• Hybrid search — RRF fusion of vector similarity (384-dim ONNX) and BM25 (FTS5), with entity boost and project boost
• Entity detection — regex-based extraction of project names, tools, and topics from queries, boosting relevant results
• Topic change detection — identifies conversation shifts and increases retrieval diversity
• Memory stack — L0 identity + L1 essential stories + L2 on-demand retrieval, budget-enforced (~900 tokens)
• Vector cache — in-memory RAM cache of embeddings for fast similarity search
• Incremental indexer — polling-based with SHA-256 delta sync, heading-aware chunking
• Knowledge graph — automatic pattern-based extraction of entities and relationships (no LLM needed)
• Credential redaction — regex-based secret sanitization before storage
• Multi-source import — Claude Code (JSONL), OpenCode (SQLite), Cursor (.mdc rules), DevClaw

Install

From GitHub Releases:

# Linux / macOS
curl -fsSL https://raw.githubusercontent.com/jholhewres/anchored/main/install/install.sh | bash

From source:

git clone https://github.com/jholhewres/anchored.git
cd anchored && make build
sudo cp bin/anchored /usr/local/bin/

First run auto-downloads the embedding model (~470MB) and creates ~/.anchored/.

Setup

Add Anchored as an MCP server to your tool:

Claude Code (.claude/settings.json):

{
  "mcpServers": {
    "anchored": {
      "command": "anchored"
    }
  }
}

Cursor / OpenCode — add the same MCP config in your tool's settings.

CLI

anchored                    Start MCP server (STDIO)
anchored serve              Start MCP server (STDIO)
anchored import [sources]   Import memories from detected sources
anchored search <query>     Search memories
anchored save <content>     Save a memory
anchored list               List memories
anchored forget <id>        Remove a memory
anchored stats              Show memory statistics
anchored identity [edit]    View or edit identity file
anchored config [show|set]  View or modify configuration
anchored init               Auto-detect tools and register MCP

Import sources: claude-code devclaw opencode cursor all

MCP Tools

Tool	Description
anchored_context	Load relevant memory for the current project
anchored_search	Search across all memories (semantic + keyword)
anchored_save	Persist a fact, decision, or preference
anchored_list	List memories by category or project
anchored_forget	Remove a memory
anchored_stats	Memory overview
kg_query	Query the knowledge graph
kg_add	Add a relationship to the knowledge graph

How it works

• Hybrid search — RRF fusion of vector similarity (ONNX, multilingual) and BM25 (FTS5), with entity boost and project boost
• Entity detection — extracts project names, tools, and topics from queries to boost relevant results
• Topic change detection — identifies conversation shifts and increases retrieval diversity
• Memory stack — L0 identity + L1 essential stories + L2 on-demand, budget-enforced
• Knowledge graph — bitemporal triples with functional predicates and alias resolution, auto-extracted from memory text
• Credential redaction — regex-based secret sanitization before storage

Storage

~/.anchored/
├── data/
│   ├── anchored.db        # SQLite (FTS5 + vector cache + knowledge graph)
│   └── onnx/              # local embedding model (~470MB)
└── config.yaml

No daemon. No ports. The binary runs on demand via MCP STDIO.

Docs

• Design — memory stack, hybrid search, knowledge graph, quantization
• Architecture — project structure and implementation details
• Embedding Model — model choice, quantization, inference pipeline
• Import Sources — how each tool's data is parsed
• Changelog — version history

License

MIT