codexfi

Persistent memory for OpenCode AI agents.

The best memory system is the one that just works. No Docker. No Python. No frontend servers. Just curl -fsSL https://codexfi.com/install | bash and you're done.

What is this?

OpenCode starts every session from scratch. No memory of past decisions, established patterns, or project context. Every session, you repeat yourself.

codexfi fixes this. After every assistant turn, key facts are automatically extracted and stored locally. On every new session, relevant memories are silently injected into context. The agent just remembers.

Install

curl -fsSL https://codexfi.com/install | bash

The installer prompts for API keys, registers the plugin, and you're done.

To verify everything is configured correctly before your first session:

bunx codexfi status

Once a session runs, proof it's working: start a second session and ask the agent about something from the first — it will know.

Prerequisites

• Bun runtime
• OpenCode AI coding agent
• One extraction API key: Anthropic (recommended), xAI, or Google
• Voyage AI API key (free tier available)

How it works

1. You code normally. No commands, no /save, nothing to learn.
2. After every turn, the plugin extracts typed facts from the conversation and stores them locally in LanceDB.
3. On every LLM call, a [MEMORY] block is rebuilt into the system prompt with project context, preferences, and semantically relevant memories.
4. When you switch topics mid-session, semantic search refreshes to surface different memories.
5. Across sessions, the agent picks up where it left off — project decisions, error fixes, your preferences, everything.

All data stays on your machine at ~/.codexfi/. Nothing is sent anywhere except the embedding and extraction APIs.

Features

• Fully automatic — saves after every turn, injects on every call
• 100% local storage — LanceDB embedded in-process
• Multi-provider extraction — Anthropic Haiku (default), xAI Grok, or Google Gemini
• Code-optimised embeddings — Voyage voyage-code-3
• Typed memory system — architecture, error-solution, preference, progress, and more
• Smart deduplication — cosine similarity prevents duplicates
• Contradiction handling — new facts automatically supersede stale ones
• Compaction-proof — memory lives in the system prompt, never lost to context truncation
• Privacy filter — wrap content in <private>...</private> to exclude from extraction
• Web dashboard — codexfi dashboard for live activity, costs, and memory browser
• CLI — codexfi list, search, stats, export, forget, status

Configuration

Config at ~/.config/opencode/codexfi.jsonc:

{
  "extractionProvider": "anthropic",  // "anthropic", "xai", or "google"
  "voyageApiKey": "pa-...",
  "anthropicApiKey": "sk-ant-...",
  "similarityThreshold": 0.45,       // retrieval cutoff (0-1)
  "maxMemories": 10                  // per scope per session
}

Agent instructions (optional)

The plugin works without this, but adding instructions to ~/.config/opencode/AGENTS.md improves agent behavior — it understands the [MEMORY] block, uses the memory tool correctly, and never announces memory operations.

See plugin/README.md for the recommended AGENTS.md snippet.

Privacy

All data stays on your machine. Outbound API calls go only to Voyage AI (embeddings) and your chosen extraction provider (one per turn). Wrap sensitive content in <private>...</private> to exclude it entirely.

More

• Plugin internals — architecture, data flow, extraction providers, development setup
• Benchmark — 94.5% on DevMemBench (200 questions, 8 categories)
• E2E tests — 13 autonomous test scenarios

License

MIT