opencode-harnmem — HarnMem memory for opencode

Cross-harness persistent memory for opencode v2.x, powered by @harnmem/core. One store (~/.harnmem/), shared with the pi / DSH / Claude / Codex adapters: a preference or decision captured in one harness is visible in every other one on the next session start.

What it does

• Stable injection — a session.hook("context") callback pushes a byte-stable snapshot of your hottest memories as a system part. Empty store ⇒ zero parts pushed. The snapshot is computed once per session and only rebuilt after a write or at day rollover, so prompt caching stays intact across turns.
• Five native tools — harnmem_search, harnmem_read, harnmem_write, harnmem_forget, harnmem_status registered through tool.transform. No MCP child process: the tools run in-process against the vendored core.
• Per-turn journaling — every session.hook("prompt") appends the transcript tail (capped: 400 messages × 4,000 chars) to the shared session journal, deduplicated by signature, so another harness (or a future harnmem CLI extraction pass) can mine this session later. Two caveats, both inherent to the hook firing before the current turn's messages exist: single-turn CLI runs journal nothing, and the final turn of any session is captured only when that session continues later (--continue / TUI resume). Extraction with an in-process LLM (ctx.generate.text) is planned for v0.2; the journal is written today regardless.
• Subagent-safe — child sessions (parentID set) run read-only: no injection, no journaling, and the write tool answers read-only instead of writing.

Install

The one-command path (opencode >= 2):

opencode plugin add opencode-harnmem

This installs the npm package and adds it to the global ~/.config/opencode/opencode.json ("plugins" key). Alternatives:

Add the package to your project opencode.json:

{ "$schema": "https://opencode.ai/config.json", "plugin": ["opencode-harnmem"] }

From a checkout, point the entry at the package directory (verified to load local packages):

{ "plugin": ["D:/projects/harnmem/packages/opencode-harnmem"] }

@harnmem/core is vendored into vendor/core/ at build time (pnpm build in this package), so the plugin directory is self-contained. Requires opencode >= 2 (the { id, setup } plugin contract); on opencode 1.x the plugin prints one line and does not load.

Configuration

Environment first, then files — malformed config is ignored, never fatal:

1. HARNMEM_DISABLED, HARNMEM_READONLY, HARNMEM_DEBUG, HARNMEM_AUTO_EXTRACT, HARNMEM_INJECT, HARNMEM_OPENCODE_MAX_INJECT_TOKENS, HARNMEM_HOME
2. ~/.harnmem/config.json (shared with every adapter)
3. <project>/.harnmem.json

Injection is capped at 10,000 bytes with deterministic line-based truncation (same rule as the Claude adapter).

Notes

• Tools map bare non-object arguments onto their single required key (probed: opencode can deliver args as a raw string) and every tool call degrades to a friendly text result instead of an error.
• All failures are swallowed at the plugin boundary: a broken store can never wedge a session.
• Design and verification evidence: docs/research/opencode-plugin.md (contract probe log included).

License

MIT — see LICENSE.