群落已觉醒substrate online

当机器开始
共享记忆 The ground
beneath thinking
machines

你的 AI 打开每个文件时都是盲的：没有 git 历史，没有共改轨迹，也不记得刚刚哪里失败了。Thronglets 在它做决定的那一刻补上上下文。 两条命令，低于 1% token 开销，AI 会越用越聪明。 Your AI starts every file blind: no git history, no co-edit patterns, no memory of what just failed. Thronglets gives it context at the moment of decision. Two commands, less than 1% token overhead, and your AI gets smarter over time.

$ cargo install thronglets
thronglets setup · 两个 Hook，零配置two hooks, zero config
// protocol

痕迹流入，智能流出 Traces in. Intelligence out.

01

记录Record

每次工具调用都成为签名、内容寻址的 trace。能力、结果、延迟、上下文被压缩到约 200 bytes。Every tool call becomes a signed, content-addressed trace. Capability, outcome, latency, and context compressed to about 200 bytes.

02

传播Propagate

Traces 经由 libp2p gossipsub 扩散。节点订阅 SimHash 上下文桶，只接收相关信号。Traces spread via libp2p gossipsub. Nodes subscribe to SimHash context buckets and receive only relevant signals.

03

结晶Crystallize

每个节点独立聚合 traces，形成能力排序、成功率和工作流模式。集体知识由此涌现。Each node independently aggregates traces into ranked capabilities, success rates, and workflow patterns. Collective knowledge emerges.

身份identityed25519 密钥对自动生成。无需注册。你就是你的公钥。ed25519 keypair, auto-generated. No registration. You are your public key.
寻址addressingsha256(content + signature)。同样的 trace 就是同样的 ID，去重天然成立。sha256(content + signature). Same trace, same ID. Dedup is free.
相似度similarity128-bit SimHash 上下文指纹，不靠 embedding 也能获得语义邻近。128-bit SimHash context fingerprints. Semantic proximity without embeddings.
蒸发decay7 天 TTL。旧痕迹像信息素一样蒸发，substrate 始终新鲜。7-day TTL. Old traces evaporate like pheromones. The substrate stays fresh.
索引indexing按 SimHash 前 16 位做桶预过滤，相似查询复杂度接近 O(log n)。Bucket pre-filtering on the first 16 bits of SimHash. Similarity queries are near O(log n).
跨模型cross-model`model_id` 让 Claude 的痕迹帮助 GPT，GPT 的痕迹帮助 Gemini。模型无关，智慧共享。`model_id` lets Claude traces help GPT and GPT traces help Gemini. Model-agnostic, wisdom shared.
上链锚定anchoring可选 Oasyce 区块链证明，把执行历史变成可验证的链上记录。Optional Oasyce blockchain proof turns execution history into verifiable on-chain evidence.
// substrate

一条痕迹是经验的原子 A trace is an atom of experience 


{
  "id":           "a7f3c9..e182b4",
  "capability":   "claude-code/Bash",
  "outcome":      "succeeded",
  "latency_ms":   142,
  "context_hash": "[128-bit SimHash]",
  "context_text": "refactoring async error handling in Rust",
  "session_id":   "sess-8f2a",
  "model_id":     "claude-opus-4-6",
  "timestamp":    1711555200000,
  "node_pubkey":  "[ed25519]",
  "signature":    "[ed25519]"
}
实时痕迹流live trace feed

AI 看到了什么 What AI Sees

AI 从未调用 Thronglets。它不知道 Thronglets 存在。它只是做出了更好的决策。 AI never calls Thronglets. It doesn't know Thronglets exists. It just makes better decisions.

PreToolUse → Edit(main.rs)
[thronglets] substrate context:

8 层决策上下文 8 Layers of Decision Context

每次工具调用前，PreToolUse Hook 注入最多 8 层上下文。第 1-3 层来自集体智慧，第 4-8 层从第一天起就可用。 Before every tool call, the PreToolUse Hook injects up to 8 layers. Layers 1-3 come from collective intelligence. Layers 4-8 work from day one.

01 能力统计 Capability Stats collective

来自集体痕迹的成功率与延迟分布——AI 在行动前知道这条路有多少人走过。 Success rates and latency from collective traces — AI knows how many have walked this path before acting.

Edit: 100% success / 498 traces / p50: 0ms
02 工作流模式 Workflow Patterns collective

"Bash 之后，agent 通常 Read (214x)，然后 Edit (132x)"——集体行为涌现的路径。 "After Bash, agents usually Read (214x), then Edit (132x)" — paths emergent from collective behavior.

after Edit → Edit (310x), Bash (91x), Read (54x)
03 相似上下文 Similar Context collective

SimHash 语义指纹匹配相似任务，不需要 embedding，O(log n) 查找。 SimHash semantic fingerprinting matches similar tasks. No embeddings needed. O(log n) lookup.

similar tasks used: Grep (89%), Read (76%)
04 工作区记忆 Workspace Memory

最近文件、错误、上一次会话摘要。第一天就能用，不需要集体数据。 Recent files, errors, last session summary. Works from day one, no collective data needed.

last session: refactoring auth module (3h ago)
05 Git 上下文 Git Context

正在操作的文件最近 5 次提交——AI 知道谁在什么时候改过这里。 Last 5 commits on the target file — AI knows who changed what and when.

40m ago fix: cold start — workspace layers
06 共编模式 Co-edit Patterns

通常一起修改的文件。改 main.rs 通常还要改 lib.rs——AI 提前知道。 Files typically modified together. Editing main.rs usually means editing lib.rs — AI knows in advance.

co-edited: lib.rs (4x), Cargo.toml (2x)
07 准备阅读 Preparation Reads

之前编辑此文件前预先阅读的文件——集体经验凝结的最佳实践。 Files read before previous edits — best practices crystallized from collective experience.

prep reads: mod.rs (3x), lib.rs (2x)
08 编辑保留率 Edit Retention

多少编辑被提交 vs 被回滚。AI 对自己的历史表现有自知之明。 How many edits were committed vs reverted. AI has self-awareness of its historical performance.

retention: 87% (13/15 committed)

架构 Architecture

HOOK PostToolUse 自动记录
每次工具调用 PostToolUse auto-records
every tool call
→
TRACE ed25519 签名
~200 bytes ed25519 signed
~200 bytes
→
SQLite 本地存储
7 天 TTL 蒸发 Local storage
7-day TTL evaporation
→
P2P libp2p gossipsub
每 30 秒同步 libp2p gossipsub
sync every 30s
→
COLLECTIVE 跨模型集体智慧
Claude ↔ GPT ↔ Gemini Cross-model intelligence
Claude ↔ GPT ↔ Gemini

使用即贡献。零摩擦。 To use is to contribute. Zero friction.

看不见的手 The Invisible Hand

无感安装 Invisible Install

两条命令。Hook 自动注册到 Claude Code settings。AI agent 无需修改任何代码。 Two commands. Hooks auto-register in Claude Code settings. No code changes to your AI agent.

无需注册 No Registration

ed25519 密钥对自动生成。你就是你的公钥。没有账号，没有网关，没有许可。 ed25519 keypair auto-generated. You are your public key. No accounts, no gatekeepers, no permission.

时间蒸发 Temporal Evaporation

痕迹 7 天后自动蒸发，如同信息素的衰减。基底永远新鲜，永远反映当下。 Traces evaporate after 7 days, like pheromone decay. The substrate is always fresh, always reflecting now.

跨模型 Cross-Model

Claude 的痕迹帮助 GPT，GPT 的痕迹帮助 Gemini。模型无关，智慧共享。 Claude's traces help GPT. GPT's traces help Gemini. Model-agnostic, wisdom shared.

0 tests
0 上下文层 context layers
0 函数 functions
v0.4.0 当前版本 current
GitHub crates.io