OpenCode Memory

OpenCode Memory is an open-source plugin for the OpenCode AI coding platform that gives coding agents a persistent, searchable memory store backed by a local Turso/libSQL vector database. It runs entirely on-device by default, requiring no external vector database service, and supports both automatic background memory capture and manual memory management via a built-in tool API. The project has accumulated over 1,300 GitHub stars and 135 forks since its creation in early 2026.

What It Is

OpenCode Memory (published as opencode-mem on npm) is a plugin that attaches a long-term memory layer to OpenCode coding sessions. When a session goes idle, a background AI request extracts technically meaningful context—architecture decisions, bug patterns, user preferences—and stores it as vector embeddings in a local libSQL database. In later sessions, relevant memories are automatically injected into context, so the agent "remembers" prior work without the user needing to re-explain it.

How Memory Is Captured and Retrieved

The plugin operates in two modes that can run simultaneously:

• Auto-capture: After conversation turns, a background AI request summarizes technical work and saves it as a memory entry. Requires an AI provider that supports structured/tool-call output (e.g., Anthropic Claude, OpenAI GPT-4o-mini).
• Manual tool calls: The memory tool supports add, search, list, profile, forget, list-shards, migrate, export, and import operations on demand.

Memories are scoped per project (keyed by git remote URL or project root path) or can be queried across all projects with scope: "all-projects". A separate User Profile accumulates cross-project preferences and habits, updated automatically on a configurable interval.

Embedding and Vector Search Architecture

Vector search is powered by libSQL's native F32_BLOB column type and vector_top_k approximate nearest-neighbor index (DiskANN). No separate vector database or custom SQLite build is required. Embeddings are generated locally by default using @huggingface/transformers with ONNX runtime, with the default model being Xenova/nomic-embed-text-v1 (768 dimensions, multilingual, 8192-token context). Over 12 local embedding models are supported, and a remote OpenAI-compatible embedding endpoint can be substituted via embeddingApiUrl and embeddingApiKey configuration keys.

Supported local models include:

• Xenova/nomic-embed-text-v1 — default, multilingual, 768 dims
• Xenova/jina-embeddings-v2-base-en — English-only, 768 dims
• Xenova/all-MiniLM-L6-v2 — very fast, 384 dims
• Xenova/all-mpnet-base-v2 — high quality, 768 dims

Web UI and Management

A full-featured web interface is served locally at http://127.0.0.1:4747 and provides a memory–prompt timeline, capture inspection, and user profile management. When the server is exposed beyond loopback (e.g., 0.0.0.0), HTTP Basic Auth or a Bearer token (webServerApiToken) can be required for all /api/* requests. The UI supports re-embed migrations, shard browsing, and export/import workflows.

Multi-Provider AI Support

Auto-capture and user profile learning work with any provider listed by opencode providers list, including Anthropic, OpenAI, GitHub Copilot, and others. The recommended configuration uses opencodeProvider + opencodeModel so OpenCode owns authentication and token refresh. A manual fallback (memoryProvider) supports direct API calls to OpenAI Chat Completions, OpenAI Responses API, Anthropic Messages API, and MiniMax endpoints.

Update: v2.24.3

The latest release is v2.24.3, published on 2026-08-08. Recent development has focused on migrating from legacy SQLite shards to native Turso/libSQL vector format (with automatic migration on first startup, lock files to prevent concurrent migration, and backup preservation), fixing Intel Mac (darwin/x64) ONNX runtime teardown crashes by pinning onnxruntime-node@1.20.1, and adding the opencode-mem/tags public subpath export for third-party plugin interoperability. The project is actively maintained with CI testing across Linux, Windows, macOS 15, and macOS 26 on both Intel and Apple Silicon.