coding-agents
vectorize-io/hindsightHindsight:会学习的 Agent 长期记忆系统,自动召回/保存、知识页、深度反思与按仓库隔离的记忆银行。
自进化团队记忆,以纯 Markdown 存于 Git:原生 Cordis 插件注册 38 个裸名记忆工具,并按每次组装动态注入 prompt-signals 段;含 RPE 强化、衰减与睡眠巩固;与 Claude Code(MCP)、OpenClaw 宿主共享同一数据仓;已对 DSH 0.1.0-rc.6 实测。
English | 中文
Co-Engram team memory for DeepSeek Harness — a native Cordis plugin.
ctx.tools with bare names (engram_search, engram_create, …) — same tool set as the Claude Code host.memory:co-engram prompt section (order 120): top tags, skill catalog, path overview and pending-proposal count are re-evaluated at every prompt assembly — write a memory and the next message already reflects it.dsh plugin --profile <name> add @co-engram/dsh
That's it — the package declares a dsh.bundle patch, so the plugin is activated as a profile layer with zero manual cordis.patch.yml editing. Point co-engram at your data repo once (shared with other hosts):
npm install -g @co-engram/claude-code # provides the `co-engram` CLI
co-engram config data-root $HOME/team-memory
- id: co-engram
name: '@co-engram/dsh'
config:
language: en # or zh — tool descriptions & prompt language
startMaintenance: true # background reinforce/forget/consolidate
proposalEnabled: true # implicit-capture proposal engine
startViewer: true # web viewer (default: follows proposalEnabled, holder-gated, port 18899)
defaultCreatedBy: '' # fallback creator (default: git author)
See docs/host-dsh.md for the full reference.
MCP bridge (dsh-mcp-client) | this plugin | |
|---|---|---|
| Tool names | mcp__co-engram__* | bare engram_* |
| Prompt guidance | server instructions are not bridged — signals lost | native dynamic section, per-assembly |
| Claude Code hooks side effect | yes (auto-installs into ~/.claude/settings.json) | no |
MIT