Skip to content

Memcore Cloud 2026.6.20

Choose a tag to compare

View all tags
@strmforge strmforge released this 20 Jun 12:44
v2026.6.20
a0f35bb

Memcore Cloud 2026.6.20

2026.6.20 is a safety-focused public release for Memcore Cloud. It lowers the
risk of local agent memory overreach, especially OpenClaw-style interception,
while keeping recall source-backed and local-first.

Memcore Cloud remains a local AI memory library: original records stay on the
user's machine, recall is source-backed, and reusable experience must keep a
traceable evidence trail.

What This Release Contains

  • Local AI memory library: original records remain the local source of
    truth, while source refs, library ids, and receipts make recall auditable.
  • Safer local-agent authority: Zhiyi is a memory and evidence layer, not a
    default replacement for the host agent. OpenClaw-style interception defaults
    to off; direct Zhiyi answers and platform actions require explicit entry or
    authorization.
  • Low-resource defaults: watchers default to a light profile, a 5-second
    interval, and narrow source selection instead of full-source 250ms scans.
  • More honest local tool display: public docs may describe supported AI
    tools, but the local console should emphasize tools actually detected on the
    current machine.
  • Pre-work checks: work_preflight / preflight doctor can run as a quick
    daily smoke path, with full diagnostics reserved for explicit troubleshooting.
  • Source-backed recall: recall remains compact by default, with source refs
    first and raw excerpts only when explicitly requested.
  • Evidence-bound model path: MiniMax, DeepSeek, and OpenAI-compatible
    models can be used for evidence-bound answer refinement and candidate
    checking without treating models as benchmark-only toys.
  • Fast model diagnostics: the model matrix can compare baseline top+pack
    two-call behavior with a single-call fast audit, including call count,
    latency, CPU/RSS, decision drift, and risk flags.
  • Benchmark diagnostics, not leaderboard claims: public docs now separate
    no-key retrieval diagnostics, internal answer judging, and official
    evaluator paths so scoring work does not masquerade as a published benchmark.
  • Evaluation guardrails: daily use, targeted regression, and offline
    benchmark entry points are separated so scoring work does not overload a
    workstation or silently become the daily path.

中文

2026.6.20 是一次以安全降险为核心的公开版本。它降低本机 agent 记忆层越权的
风险，尤其是 OpenClaw 这类拦截入口默认抢答的问题，同时继续保持本机优先和
可回源召回。

本版本包含

  • 更安全的本机 agent 权限边界：知意是记忆和证据层，不默认替代宿主
    agent。OpenClaw 这类拦截入口默认关闭；直接回答和平台动作必须有显式入口
    或授权。
  • 低资源默认：watcher 默认 light profile、5 秒间隔、窄来源选择，不再默认
    全来源 250ms 扫描。
  • 更诚实的本机工具展示：公开文档可以说明支持哪些 AI 工具，但本机控制台应
    优先展示当前机器真实检测到的工具。
  • 开工前检查：work_preflight / preflight doctor 支持日常快速 smoke，
    full 体检保留给显式排障。
  • 可回源召回：默认返回紧凑 source refs；只有明确需要原文证据时才取 raw
    excerpt。
  • 证据绑定模型路径：MiniMax、DeepSeek 和 OpenAI-compatible 模型可以用于
    证据绑定回答、候选判断和经验精炼，不再只作为 benchmark 工具。
  • fast 模型诊断：模型矩阵可比较 top+pack 双调用 baseline 和单调用 fast
    audit，记录调用数、耗时、CPU/RSS、逐题漂移和风险标记。
  • 评分诊断不冒充榜单：公开文档已区分免费检索诊断、内部答案判分和官方
    evaluator 路径，避免把本地诊断说成公开 benchmark 成绩。
  • 评测护栏：日常入口、定向回归和离线 benchmark 分开，避免跑分任务拖垮
    工作机，也避免把评分路径误当成日常路径。

Boundaries

  • Local benchmark and model-matrix reports are diagnostics, not official
    leaderboard scores.
  • Benchmark run outputs, caches, and local R730XD pressure-test artifacts are
    not part of the public release payload.
  • Direct Zhiyi answers and platform actions remain explicit-entry or
    explicit-authorization paths; ordinary OpenClaw chat is passive by default.
Assets 8
Loading