Anchor

Cross-agent memory for AI coding agents. Local-first.

Switch agents — Claude Code, Codex, Cursor, Cline, Antigravity — and your project context comes with you.

The Problem

Developers build deep context with an AI agent—defining project preferences, structural decisions, and mental models—over several hours of work. However, when switching tools due to usage limits, collaboration needs, or specialized features, the new agent starts cold. You are forced to re-explain choices and answer questions that were already settled.

This is not a model limitation, but a portability problem: session memory is locked inside each individual AI tool or provider's silo.

The Solution

Anchor is a local-first Model Context Protocol (MCP) server and an open Agent Skill. It enables any modern AI coding agent to read from and write to a single, durable, project-specific memory.

When a new agent session starts, it retrieves a highly compressed, token-budgeted gist of what matters—rather than a full, noisy conversation transcript. No external accounts or API keys are required. All data is stored locally in a SQLite file at ~/.anchor/memory.db.

Visual Architecture

The diagram below shows how Anchor sits between your agents and your local database. Any agent that speaks the Model Context Protocol (MCP) or Vercel Skills specification can automatically read from and write to your shared local memory.

flowchart TB
    subgraph Agents
        A1[Claude Code]
        A2[Codex CLI]
        A3[Cursor]
        A4[Antigravity]
        A5[Cline / Continue / ...]
    end

    subgraph "Anchor MCP Server (Node 20+, TypeScript)"
        T[memory_recall · memory_remember · memory_supersede · memory_forget · memory_list · memory_summary · memory_context]
        R[Retrieval: BM25 FTS5 + optional vector hybrid → typed rerank → token-budgeted gist]
        C[Capture: secret redaction · injection scrub · scope isolation]
        T --> R
        T --> C
    end

    subgraph "Local storage"
        DB[(SQLite at<br/>~/.anchor/memory.db)]
    end

    subgraph "Optional: embeddings"
        E1[Ollama local]
        E2[OpenAI]
        E3[Gemini]
        E4[Voyage]
    end

    A1 -->|MCP stdio| T
    A2 -->|MCP stdio| T
    A3 -->|MCP stdio| T
    A4 -->|MCP stdio| T
    A5 -->|MCP stdio| T
    R -->|read/write| DB
    C -->|write| DB
    R -.->|hybrid recall| E1
    R -.->|hybrid recall| E2
    R -.->|hybrid recall| E3
    R -.->|hybrid recall| E4

    style T fill:#11161d,stroke:#06b6d4,color:#e6edf3
    style R fill:#11161d,stroke:#06b6d4,color:#e6edf3
    style C fill:#11161d,stroke:#06b6d4,color:#e6edf3
    style DB fill:#0a0e14,stroke:#06b6d4,color:#e6edf3

Installation & Setup

Set up Anchor and connect it to your favorite AI coding agents in under a minute.

1. Install, Initialize & Auto-Register

Run the following command to install the CLI, initialize your local environment, and automatically register the MCP server inside all detected local AI coding tools (Claude Code, Cursor, Codex, Windsurf, Cline, and Antigravity):

npx @anchormem/anchor init

That's it! Running init scans your machine, locates configuration folders for your installed agents, and injects the Anchor MCP server entry directly. No manual config editing is required.

2. Share with your Team (Optional but Recommended)

To make context sharing seamless for your team, run the setup command inside your project's repository root:

npx @anchormem/anchor setup

This generates project-scoped configuration files (like .mcp.json for Claude Code, .cursor/mcp.json for Cursor, and .agents/mcp_config.json for Antigravity). Teammates who clone the repository will have their AI tools auto-detect and load Anchor without any manual setup!

3. Verification & Manual Fallbacks

Run the built-in diagnostic suite to confirm everything is configured perfectly:

npx @anchormem/anchor doctor

Should you ever need to perform manual registration, Anchor supports standard MCP configurations. Simply map the server key anchor to the command anchor-server (or npx -y @anchormem/anchor@latest anchor-server) in your tool's settings.

How It Works

Four-Tier Memory Structure

Anchor stores project context in four distinct, typed structures, preventing unstructured context drift.

Type	Description	Example
Fact	A durable developer preference or architectural constraint.	Uses pnpm, not npm
Decision	An active choice made during a task, along with its specific rationale.	Use PostgreSQL for order service due to ACID compliance
Episode	A short, 1-3 sentence summary of a task, automatically recorded by the agent.	Added rate limiting via Redis token bucket; updated src/auth/middleware.ts
Artifact	A specific file pointer, symbol, or reference URL.	`src/auth/middleware.ts:42`

When a fact or decision is updated, the agent uses memory_supersede to retire the older record rather than leaving conflicting entries. Older task episodes decay in salience over time, keeping your active context clean.

Per-Language Prioritization & Boosting

In multi-language repositories (monorepos or polyglot codebases), memories are automatically scoped by language:

Auto-Tagging on Capture: When an agent saves a fact, decision, or episode, Anchor automatically infers the target programming language (e.g., typescript, go, rust, python) by analyzing file paths, extensions, markdown code blocks, or syntax patterns.
Context-Aware Boosting: When an agent requests context, Anchor detects the active programming language from the agent's current files. Memories matching that language receive a 1.5x rank weight boost and a +0.5 base score.
No Hidden Context: Cross-language/general memories (like pnpm commands or architectural designs) still surface, but out-of-scope specific languages are gently deprioritized to preserve the limited context token budget.

Performance & Context Optimization

Anchor uses hierarchical structuring and a highly optimized local retrieval engine to keep your agent's context small and relevant.

Measured offline against pasting full prior transcripts (the realistic alternative for manual context sharing), Anchor achieves a 44x context compression ratio with 100% recall of required facts.

Benchmark Results (8 Scenarios)

The following evaluations are fully reproducible using the local harness (node tests/eval/run.mjs). They cover diverse project domains including backend development, SaaS billing, ML pipelines, infrastructure, and mobile apps.

Scenario	Category	Baseline (Cold)	With Anchor (Warm)	Context Saved	Hit Rate	FTS Precision
Auth rate limiting	Backend API	9,400 tokens	224 tokens	97.6%	100% (5/5)	80.0%
Stripe billing migration	SaaS Payments	7,800 tokens	163 tokens	97.9%	100% (3/3)	100.0%
Turborepo monorepo	Infrastructure	12,500 tokens	304 tokens	97.6%	100% (6/6)	83.3%
ML pipeline debugging	ML/Data	15,000 tokens	334 tokens	97.8%	100% (6/6)	83.3%
React Native state management	Mobile	11,200 tokens	271 tokens	97.6%	100% (5/5)	100.0%
EC2 to Kubernetes migration	DevOps	13,800 tokens	330 tokens	97.6%	100% (6/6)	66.7%
Security audit remediation	Security	10,500 tokens	276 tokens	97.4%	100% (6/6)	66.7%
MySQL to PostgreSQL database migration	Database	18,000 tokens	337 tokens	98.1%	100% (6/6)	83.3%
Aggregate Summary	All Domains	98,200 tokens	2,239 tokens	97.7%	100% (43/43)	82.9%

Note: In all scenarios, information leak prevention successfully validated that no superseded or redacted content was retrieved (0 leaks).

Core Product Features

Scope-Isolated Contexts: Anchor automatically detects directory boundaries based on your working directory, keeping memories isolated to the relevant project.
Per-Language Prioritization: In monorepos or multi-language projects, Anchor automatically tags memories with their active programming language and applies a 1.5x soft-boosting relevance adjustment when working inside those languages, ensuring relevant context fits inside the token budget first.
Durable Four-Tier Memory: Organizes information into distinct, structured types (Facts, Decisions, Task Episodes, and Artifacts) for structured and precise recall.
Local SQLite Engine: Uses local SQLite FTS5 (BM25) for ultra-fast, zero-dependency, and deterministic retrieval.
Embedded Security & Redaction: Automatically scrubs credentials (such as Stripe, Slack, AWS, and OpenAI keys) and prompt-injection keywords at write-time, before they ever hit the disk.
Hybrid Semantic Search (Optional): Integrates seamlessly with Ollama (local), OpenAI, Gemini, or Voyage to provide keyword-independent vector-hybrid retrieval.

Advanced Configuration & Commands

Semantic Recall via Vector Embeddings

Anchor's default retrieval is BM25 over SQLite FTS5—fast, deterministic, and requiring no external setup. For keyword-independent recall (e.g., retrieving "token verification key rotation" when the query is "auth key renewal"), you can enable semantic vector search.

Supported Providers

Provider	Default Model	Best For
Ollama	`nomic-embed-text`	Completely local, free, offline retrieval
OpenAI	`text-embedding-3-small`	Highly optimized cloud-hosted retrieval
Gemini	`text-embedding-004`	Google Cloud ecosystems
Voyage	`voyage-3`	Anthropic-recommended embeddings

Setup Examples

Ollama (Local)

ollama pull nomic-embed-text
export ANCHOR_EMBED_PROVIDER=ollama

OpenAI

export ANCHOR_EMBED_PROVIDER=openai
export ANCHOR_OPENAI_API_KEY=sk-your-key-here

Gemini

export ANCHOR_EMBED_PROVIDER=gemini
export ANCHOR_GEMINI_API_KEY=your-key-here

Voyage

export ANCHOR_EMBED_PROVIDER=voyage
export ANCHOR_VOYAGE_API_KEY=your-key-here

Backfilling Embeddings

For existing scopes, backfill vector representations by running:

anchor reembed

Auto-load Memory on Session Start (Zero Tool Calls)

You can automatically inject your project's history the moment an agent session starts, without requiring the agent to make manual tool calls. Anchor reads your directory scope, compiles a 1,500-token gist, and inserts it directly.

For Claude Code, add the following hook configuration to your ~/.claude/settings.json:

{
  "hooks": {
    "SessionStart": [{ "command": "anchor hook claude-code session-start" }]
  }
}

For other agents, pipe the scope payload directly into the agent's initialization or system prompt argument:

echo '{"cwd":"'"$PWD"'"}' | anchor hook generic session-start

If the current directory contains no stored memories, the hook returns nothing, preserving your token budget.

HTTP Transport Mode

By default, Anchor communicates via MCP over standard I/O (stdio). For browser-based agent interfaces, containerized pipelines, or remote development, Anchor can be run in HTTP transport mode:

# Enable HTTP server via flag
anchor-server --http

# Or via environment variable
ANCHOR_TRANSPORT=http anchor-server

# Custom port configuration (defaults to 3838)
ANCHOR_HTTP_PORT=4000 ANCHOR_TRANSPORT=http anchor-server

[!CAUTION] HTTP transport binds exclusively to 127.0.0.1 and enforces strict security headers, CORS restrictions to localhost origins, rate limiting (100 requests/minute), and host-header DNS rebinding protection. Never expose this server directly to public networks.

Interactive CLI Console & Command Reference

Anchor includes an interactive console for manual memory management and search. Open it anytime by running:

anchor

Within the interactive shell, type a term to search, or use commands like /remember, /supersede, or /forget.

CLI Command Reference

Command	Purpose
`anchor init`	Sets up the `~/.anchor` environment database.
`anchor status`	Displays active database location and memory counts.
`anchor list`	Lists all memories recorded within the current directory scope.
`anchor diff --since 1d`	Displays changes made to the database (supports intervals like `1h`, `7d`, `30d`).
`anchor replay`	Provides a chronological breakdown of decisions and episodes.
`anchor doctor`	Performs diagnostic health checks on directories, scopes, and configurations.
`anchor export > backup.json`	Backs up all memories to a JSON file.
`anchor import backup.json`	Restores database states from a JSON backup.
`anchor prune`	Automatically archives or drops older, low-salience task episodes.

</details>

---

## Security & Privacy

Anchor is designed for enterprise grade security and private-first operations:
- **Client-Side Redaction:** Scans and redacts standard credential formats (AWS keys, Slack webhooks, OpenAI/Stripe API keys, and JWTs) before saving memories to disk.
- **Prompt Injection Defense:** Filters out common injection payloads (such as "ignore previous instructions") at write-time.
- **Local Isolation:** Database files are kept with strict POSIX permissions (`0600` for database files, `0700` for directories). No telemetry is captured, and no outbound network calls are made.

## Development Status

Phase 0 and Phase 1 are complete. Current work focuses on polishing distribution and platform tooling:
- Durable SQLite schemas, BM25 indexing, and vector-hybrid support are fully implemented.
- Automatic redaction, scope mapping, and local permission hardening are in place.
- Hook frameworks for Claude Code, Antigravity, and other generic CLI interfaces are functional.
- The evaluation harness passes all integration scenarios.

## Contributing

We welcome contributions to extend the Anchor ecosystem. Focus areas include:
- Verification and installation scripts for new development environments.
- Additional regex patterns for secret redaction.
- Optimization contributions to the BM25 reranking engine.

Please refer to [CONTRIBUTING.md](CONTRIBUTING.md) for style guidelines and repository structures.

## License

This project is licensed under the [MIT License](LICENSE).

Name		Name	Last commit message	Last commit date
.agents		.agents
.anchor		.anchor
.github/workflows		.github/workflows
packages		packages
site		site
tests		tests
.gitignore		.gitignore
.mcp.json		.mcp.json
Agent.md		Agent.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
PUBLISHING.md		PUBLISHING.md
README.md		README.md
SECURITY.md		SECURITY.md
package-lock.json		package-lock.json
package.json		package.json
pnpm-workspace.yaml		pnpm-workspace.yaml
tsconfig.base.json		tsconfig.base.json

Navigation Menu

Repository files navigation

Anchor

The Problem

The Solution

Visual Architecture

Installation & Setup

1. Install, Initialize & Auto-Register

2. Share with your Team (Optional but Recommended)

3. Verification & Manual Fallbacks

How It Works

Four-Tier Memory Structure

Per-Language Prioritization & Boosting

Performance & Context Optimization

Benchmark Results (8 Scenarios)

Core Product Features

Advanced Configuration & Commands

Supported Providers

Setup Examples

Ollama (Local)

OpenAI

Gemini

Voyage

Backfilling Embeddings

CLI Command Reference

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

Anchor

The Problem

The Solution

Visual Architecture

Installation & Setup

1. Install, Initialize & Auto-Register

2. Share with your Team (Optional but Recommended)

3. Verification & Manual Fallbacks

How It Works

Four-Tier Memory Structure

Per-Language Prioritization & Boosting

Performance & Context Optimization

Benchmark Results (8 Scenarios)

Core Product Features

Advanced Configuration & Commands

Supported Providers

Setup Examples

Ollama (Local)

OpenAI

Gemini

Voyage

Backfilling Embeddings

CLI Command Reference

About

Resources

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages