opencode-plugin-memory

Project-scoped persistent memory for OpenCode. The LLM autonomously saves, updates, and deletes knowledge across sessions — no background pipeline, no cloud dependency.

How it works

• Four tools (memory_add, memory_update, memory_delete, memory_read) are registered with OpenCode and available to the LLM during every session
• A compact summary of all active memories is injected into every system prompt automatically
• Memory is stored as plain Markdown in <projectDir>/.opencode/memory/MEMORY.md — human-readable and inspectable at any time

Architecture: Claude Code–style hot-path. The LLM decides what to remember and when. There is no separate consolidation model or session-end pipeline.

Installation

Install the published package:

npm i @chncaesar/opencode-plugin-memory

Then add it to your OpenCode configuration:

// opencode.json or ~/.config/opencode/opencode.jsonc// Option 1 — from npm"plugin": ["@chncaesar/opencode-plugin-memory"]
// Option 2 — from a local clone"plugin": ["/path/to/opencode-plugin-memory"]
// Option 3 — with custom token budget"plugin": [
["@chncaesar/opencode-plugin-memory", { "maxSummaryChars": 3000 }]
]

Configuration

Option	Type	Default	Description
maxSummaryChars	number	2000	Max characters used by the memory-entry summary. The fixed writing policy is added separately.
enableLog	boolean	true	Write an operation log to .opencode/memory/plugin.log. Set false to disable.

// Disable logging"plugin": [
["@chncaesar/opencode-plugin-memory", { "enableLog": false }]
]

Storage

The plugin creates .opencode/memory/ automatically on first write — no manual setup needed.

<projectDir>/
.opencode/
memory/
MEMORY.md # main store — LLM writes here, you can edit too
memory_summary.md # compact index — auto-generated, do not edit
plugin.log # operation log (add/update/delete + errors), if enableLog=true

Each entry looks like:

### [MEM-001] Store monetary amounts as integer cents- added: 2026-08-26
- tags: coding-style
Always store monetary amounts as integer cents, not floating-point values. Parse input with CurrencyUtils.textToCents().

Deleted entries are moved to an ## Archived section rather than erased, so you can review or restore them by editing the file directly.

When does the LLM save a memory?

The system prompt and tool description guide the LLM to call memory_add when:

• The user says "remember this" / "save this"
• A reusable coding rule or project convention is discovered
• A recurring bug pattern is fixed and the lesson should persist
• The user corrects how the LLM communicates, reasons, edits, tests, or collaborates
• The user confirms that a non-obvious approach worked well and should be repeated

Unscoped corrections and preferences are treated as durable by default; the user does not need to say "remember this". Instructions explicitly limited to the current answer, task, file, or one-time situation are not saved. The same complete policy and few-shot examples are injected whether the memory store is empty or already contains entries.

Development

npm install
npm run typecheck # type-check only, no emit
npm run build # compile to dist/
node --test # run unit tests (Node built-in runner)

Design notes

• No global memory — scoped to the project directory OpenCode runs in
• No vector search — retrieval is keyword-based or full-list; the summary handles ambient recall
• No concurrent write protection — designed for single-user usage; do not add locking without tests
• Zod version — uses tool.schema from @opencode-ai/plugin (zod 4), never import { z } from "zod" directly