opencode-memelord

OpenCode plugin for memelord -- persistent memory for coding agents.

What it does

Gives your OpenCode agent persistent memory that improves over time. The plugin provides everything out of the box:

Memory tools (replaces the MCP server):

| Tool | Purpose | | ------------------- | ----------------------------------------------------------------------- | | memory_start_task | Retrieve relevant memories via vector search at the start of every task | | memory_report | Store corrections, user inputs, or codebase insights | | memory_end_task | Rate retrieved memories and record task outcome | | memory_contradict | Flag an incorrect memory and delete it | | memory_status | Show memory system stats |

Lifecycle hooks (automatic, no agent action needed):

| OpenCode event | Purpose | | -------------------- | ------------------------------------------------------- | | session.created | Inject top memories into context | | tool.execute.after | Record tool failures for pattern detection | | session.idle | Analyze transcript for self-corrections and discoveries | | session.deleted | Embed pending memories, run weight decay |

Install

Add to your global OpenCode config (~/.config/opencode/opencode.json):

{
  "plugin": ["opencode-memelord@latest"]
}

That's it. OpenCode auto-installs the plugin and all dependencies at startup.

How it works

• Global database -- memories are stored at ~/.config/memelord/projects/<project>/memory.db, keyed by git remote URL. Multiple worktrees of the same repo share the same database.
• Local embeddings -- uses Xenova/all-MiniLM-L6-v2 (384-dim, quantized, runs on CPU) via @huggingface/transformers. No API keys needed. The model is lazy-loaded on first use.
• Uses the memelord SDK directly -- same memory lifecycle, scoring, and decay algorithms. Same analysis logic for detecting self-corrections, discoveries, and failure patterns.

Memory lifecycle

1. Session starts -- top memories by weight are injected into context
2. Agent works -- tool failures are tracked automatically
3. Agent finishes responding -- transcript is analyzed for self-corrections (failed tool -> same tool succeeds with different input) and discoveries (high-token exploration sessions)
4. Session ends -- new memories are embedded and weight decay runs

Memories that consistently help survive. Memories that don't get garbage collected over time.

Requirements

• OpenCode v1.0+

License

MIT