Skip to content

Releases: realchendahuang/opencode-hermes-memory

v0.3.1

Choose a tag to compare

@realchendahuang realchendahuang released this 11 Aug 16:03

内部清理与修正,无用户可见行为变化。

Fixed

  • consolidate 冷却时间不再被"条目太少"的无操作检查消耗(<2 条直接返回,不占 24h 冷却)

Changed

  • 删除死代码链:`formatForSystemPrompt` 及其独占的 snapshot/renderBlock/failure-injection 机器(净减 79 行)
  • 回归测试 21 → 25 断言(新增 `extractOperations` 解析容错覆盖)

安装:`opencode plugin github:realchendahuang/opencode-hermes-memory`

v0.3.0

Choose a tag to compare

@realchendahuang realchendahuang released this 11 Aug 15:45

Fixed

  • `touchEntry` 刷新 `last=` 时丢失 `supersedes`/`superseded` 演化链元数据
  • 首次安装时日志目录不存在导致日志静默失败——初始化时自动 `mkdir`
  • 内部审查会话的消息会触发纠正检测 / 记忆自动注入 / 轮次计数——通过内部会话 ID 集合同步跳过
  • 删除 `session.idle` 事件处理中重复的 `lastIdleSession` 检查

Added

  • 后台审查 prompt 注入已有记忆清单(≤4000 字符),避免重复保存已知事实
  • `session.deleted` 事件清理会话级状态,防 Map 随会话数无限增长
  • 回归测试 18 → 21 断言

Changed

  • 记忆自动注入块复用 `store.fenceBlock`

安装:`opencode plugin github:realchendahuang/opencode-hermes-memory`

v0.2.0 — GitHub direct install + package polish

Choose a tag to compare

@realchendahuang realchendahuang released this 11 Aug 15:14

What's new in v0.2.0

🚀 GitHub direct install

No npm needed — install straight from GitHub:

opencode plugin -g github:realchendahuang/opencode-hermes-memory

📦 Package polish

  • exports[\"./server\"] entry point (OpenCode loader resolves it first)
  • repository / bugs / homepage / author metadata
  • engines.opencode compatibility declaration
  • CI now uses bun install --frozen-lockfile for reproducible builds
  • README badge: hardcoded test count → live CI badge

🐛 Type fixes

  • Fixed type errors against @opencode-ai/plugin SDK (Message.summary object shape, Part structure, Target union narrowing)
  • tsc --noEmit clean, 18/18 regression tests green

v0.1.0 — First open-source release

Choose a tag to compare

@realchendahuang realchendahuang released this 11 Aug 14:07

🎉 First open-source release

Layered persistent memory plugin for OpenCode, ported from the Hermes agent.

Highlights

  • Layered memory: L0 standing instructions / L1 Markdown truth source / L2 token-scored retrieval
  • 5 memory tools: memory_search, memory_add, memory_replace, memory_remove, memory_history
  • Learning loop: background review on idle, flush review before compaction, auto-consolidation at capacity
  • Auto-injection: relevant memories injected per turn (score ≥ 0.4, max 2/turn)
  • Error prefetch: bash failures auto-inject related past lessons
  • Bi-temporal evolution: superseded entries kept in history.md
  • Zero runtime deps beyond @opencode-ai/plugin — pure Markdown, no vector DB

Quality

  • 18/18 regression tests green, tsc --noEmit clean
  • CI (GitHub Actions): test + typecheck on every push/PR
  • Bilingual README (EN/zh-CN), CONTRIBUTING, SECURITY, CODE_OF_CONDUCT

Install

See README.