Skip to content
BranchesTags

Repository files navigation

InfinityContext

Open-Source Context Compression & Memory Optimization for AI Agents — DeepSeek Harness (dsh), Claude Code, OpenClaw, Cursor, Dify, Ollama and any Agent Skills host

English | 简体中文 | 繁體中文 | 日本語 | 한국어 | Español | Português | Français | Deutsch | Русский

⚡ Why InfinityContext / 为什么需要它

EN — Long sessions hit the context limit. Ordinary compression throws the details away: the agent forgets what you said, loses the goal, and repeats work. InfinityContext keeps the complete record in a local SQLite + FTS5 archive while the context window stays small — so when something from hundreds of turns ago matters again, the agent finds it and returns the exact message. Offline, on your machine.

中文 — 长会话迟早撞上上下文上限。普通压缩把细节直接扔掉：智能体忘了你说过什么、丢了目标、重复干过的活。InfinityContext 把完整记录放进本地 SQLite + FTS5 归档，而上下文窗口保持精简——几百轮之前的事再被提起时，智能体能找到并取回那一句原话。离线、在本机。

What you get / 你得到什么 Why it matters / 意味着什么
🧠 Nothing is forgotten / 细节不会被忘掉 Every turn is archived and searchable down to the single message — not just a summary.
🎯 The goal never gets lost / 目标不会丢 Goals, decisions and open tasks stay retrievable after compression, so the agent does not drift or redo work.
🔎 Knows when and how to recall / 知道何时如何回忆 The skill teaches the retrieval pattern — which script, which query, which scope — instead of leaving the agent to guess.
🪶 Small window, long session / 小窗口，长会话 Keep the window lean for deep reasoning; the archive carries the volume.
🔒 Local, private, bounded / 本地、私密、有边界 No network, no telemetry, no cloud. Owner-only permissions, 30-day default retention, one-flag off switch.
⚡ Runs everywhere / 到处都能跑 DeepSeek Harness (dsh), OpenClaw, Claude Code, Cursor, Dify, Ollama — Python 3.9+ standard library only.

Before / after a compaction / 压缩前后对比

Without InfinityContext With InfinityContext
After compaction details gone, goal fuzzy, work repeated window small, details archived, goal intact
Recalling turn #12 from hours ago impossible one FTS5 query
Where your conversation lives only in the window only on your machine

Scope, stated plainly / 范围说明：details stay retrievable inside the retention window — 30 days by default, 1..3650 configurable, unbounded only with an explicit flag.

Which package should I install? / 该装哪个包？

ClawHub / registry artifact GitHub repository
What it is the security-audited portable core the complete, unabridged project
Files 15 31
OpenClaw hook, watchdog, Windows helpers (openclaw/) — ✅
Best for dsh, Claude Code, Cursor, Dify, Ollama, any SKILL.md host OpenClaw on Windows with automation
Version the audited release the newest release — integration fixes land here first

The registry package is deliberately slimmed to the auditable core — that is exactly what the security scan reviews, and nothing outside it is executed. The GitHub repository is the complete project: the identical core plus the optional host integration, so no capability is missing when you need it.

注册表包刻意精简为可审计核心（安全扫描审查的就是它，它之外没有任何代码会被执行）；GitHub 仓库是完整项目：核心完全一致，外加可选的宿主集成，需要时功能一个不少。

⚠️ Security & Privacy Disclosure — Intended Behavior / 安全与隐私披露（预期行为）

EN — InfinityContext is a local persistent store and lifecycle manager for agent sessions. By design it performs these local operations:

Operation Tool Scope control
Persist redacted conversation chunks in a local SQLite/FTS5 archive session_to_sqlite.py owner-only directory (0700/0600, or a protected DACL), fail-closed
Permanently delete expired archive files cleanup.py verified .infinity-context-archive marker + full-filename allowlist + non-recursive + --apply --confirm-destructive
Rewrite a file in place (redaction) session_to_sqlite.py --redact-file requires --allow-dir; the symlink-resolved path must stay inside it
Move the archive when the path is not ASCII-safe session_to_sqlite.py refused unless --allow-dir-fallback is given

Redaction is best-effort and the archive still holds a detailed record of your sessions. Nothing is sent anywhere — no cloud sync, no telemetry, no outbound network. Keep the archive out of synced or shared folders and run cleanup deliberately. Installing this skill means accepting these local persistence and file-mutation capabilities.

中文 — 本插件是本地持久化存储与生命周期管理器，按设计会执行以下本地操作：

操作 工具 范围控制
把脱敏后的对话片段持久化到本地 SQLite/FTS5 session_to_sqlite.py 目录 0700 / 文件 0600（Windows 为受保护 DACL），Fail-Closed
永久删除过期归档文件 cleanup.py 必须存在 .infinity-context-archive 归档标记 + 完整文件名白名单 + 不递归 + --apply --confirm-destructive
就地改写文件（脱敏） session_to_sqlite.py --redact-file 必须指定 --allow-dir，解析符号链接后仍须落在该目录内
路径非 ASCII 时改存其它目录 session_to_sqlite.py 默认拒绝，需显式 --allow-dir-fallback

脱敏是尽力而为，归档仍保留会话的详细记录；不联网、不上传、无遥测。请把归档放在 非同步、非共享目录，并谨慎执行清理。安装即表示你接受上述本地持久化与文件修改能力。

Permissions / 权限声明

Declared capability scope (mirrors the Agent Skills allowed-tools field). This is the whole surface of the published package:

Capability Used for
File read Reading the session transcript (JSONL) you pass in, and the archive it created
File write Writing the SQLite archive, its WAL sidecars, the archive marker, and — with --redact-file — rewriting one file inside a declared --allow-dir
Environment Reading INFINITY_CONTEXT_HOME, INFINITY_CONTEXT_REDACT_RULES and INFINITY_CONTEXT_NO_ARCHIVE

Not declared because not used: network, MCP, shell, subprocesses. The core never starts another program, and nothing leaves the machine.

Package boundary. The published artifact is exactly the files listed in checksums.txt (SKILL.md, README.md, 说明.md, CHANGELOG.md, SPONSORS.md, LICENSE, checksums.txt, scripts/*.py, references/*.md, sponsors/*) — no JavaScript and no PowerShell. Any optional OpenClaw/Windows integration lives in the GitHub repository under openclaw/, is not part of this package, and carries its own documentation and openclaw/checksums.txt.

OpenClaw patches (openclaw/patches/). Two optional post-install patches for the memory-tencentdb plugin — not part of the portable core.

Patch Purpose
l1-model-chain Replace hardcoded model in L1/L2/L3 extraction with ordered fallback chain + API filter
internal-session-archive Auto-archive done memory-* sessions; keep failed visible

Each patch has its own SKILL.md, self-test (--selftest), and PowerShell wrapper. Patches modify the installed plugin (not the portable core) and require a gateway restart after application.

English

What is this?

InfinityContext is a portable AI-agent skill (a standard SKILL.md bundle) that keeps long sessions usable: it compresses context, archives redacted conversation chunks into a local SQLite/FTS5 store, and retrieves exact details from an earlier turn. It runs on any Agent Skills host — first-class support for DeepSeek Harness (dsh) (drop it into .agents/skills/ and the harness discovers it) — plus Claude Code, OpenClaw, Cursor, Dify, Ollama and custom agents.

Scope, stated plainly. This package is the portable Python core. It reads a session transcript, writes a local archive, and answers searches. It has no network access, no shell, no subprocesses and no background service. Everything it can do to your files is listed in the disclosure table at the top of SKILL.md.

What it does

Step Script What happens
Archive scripts/session_to_sqlite.py transcript (JSONL) → redacted chunks in a local SQLite DB with an FTS5 index
Retrieve scripts/search.py read-only FTS5 search (trigram, with LIKE fallback for short CJK queries)
Prune scripts/cleanup.py retention cleanup of archive files; dry-run by default
Protect scripts/secure_fs.py owner-only permissions for the archive directory, database and WAL sidecars

File operations this package performs (full detail in the disclosure table at the top of SKILL.md):

Operation Tool Guard
Write redacted chunks into a local archive session_to_sqlite.py owner-only permissions, fail-closed
Permanently delete expired archive files cleanup.py archive marker + full-filename allowlist + non-recursive + --apply --confirm-destructive
Rewrite a file in place (redaction) session_to_sqlite.py --redact-file requires --allow-dir; symlink-resolved path must stay inside it
Move the archive when the path is not ASCII-safe session_to_sqlite.py refused unless --allow-dir-fallback is given

Retention policy (read this once)

Setting Value Notes
Default retention 30 days chunks older than 30 days are purged on every run, in the same transaction as the insert
Configurable range 1..3650 days --retention-days N
No time limit opt-in only --retention-days 0 is rejected unless --allow-unbounded-retention is also given
Manual pass --purge-only --output-dir <dir> applies the same policy to existing archives, without ingesting
Off switch INFINITY_CONTEXT_NO_ARCHIVE=1 the archiver writes nothing and returns status: disabled

Retention is enforced by code, not by documentation: every run deletes expired chunks from session_chunks and its FTS mirror and reports the count as purged_chunks.

Quick start (5 minutes)

Prerequisite: Python 3.9+. No Node.js, no build step, no network access.

# 1. Install (registry — a scanned artifact)
clawhub install infinitycontext --workdir ~/.agents --dir skills   # dsh / Claude Code
clawhub install infinitycontext --workdir ~/.openclaw --dir skills  # OpenClaw

# 2. Archive a transcript
python3 scripts/session_to_sqlite.py \
  --session-key my-session \
  --session-file ~/path/to/events.jsonl \
  --output-dir ~/.infinity-context/archive

# 3. Retrieve a detail later (read-only)
python3 scripts/search.py --db ~/.infinity-context/archive/my-session-*.db --query "deployment token"

# 4. Prune archive files (dry-run first)
python3 scripts/cleanup.py --archive-dir ~/.infinity-context/archive --dry-run

From source — pin the reviewed tag and verify every byte:

git clone https://github.com/Pondsi/infinitycontext.git
cd infinitycontext
git checkout --detach v1.8.9
grep -q '^version: "1.8.9"' SKILL.md || { echo "tag/version mismatch - stop"; exit 1; }
sha256sum -c checksums.txt          # macOS: shasum -a 256 -c checksums.txt

Then place the verified files into your skills directory exactly as listed in references/architecture.md — the list is explicit, so no wildcard and no cp -r is ever needed.

Where to install

Host Location Notes
DeepSeek Harness (dsh) ~/.agents/skills/infinity-context/ or <project>/.agents/skills/infinity-context/ a skill is one level deep (<root>/<dir>/SKILL.md); nested **/SKILL.md is deliberately not discovered
OpenClaw <workspace>/skills/ (highest precedence) or ~/.openclaw/skills/ discovered automatically
Claude Code ~/.claude/skills/infinity-context/ same layout
Cursor / Dify / Ollama / custom point the agent at this folder pure Python standard library

Requirements

  • Any host that loads a standard SKILL.md: dsh, Claude Code, OpenClaw, Cursor, Dify, Ollama or a custom agent.
  • Python 3.9+ only. The core contains no JavaScript, no PowerShell, no subprocesses and no network code.
  • An optional OpenClaw/Windows integration lives in the repository under openclaw/ and is not part of this package.

FAQ

Where is my data? In the directory passed as --output-dir (default ~/.infinity-context/archive): a SQLite database plus -wal/-shm sidecars, owner-only permissions.

How do I stop archiving completely? Set INFINITY_CONTEXT_NO_ARCHIVE=1; the archiver writes nothing and returns status: disabled.

How do I keep less history? --retention-days 7 — expired chunks are deleted on the next run. --purge-only --output-dir <dir> applies it immediately without ingesting anything.

Is anything sent anywhere? No. There is no network code in the package.

What about secrets in the transcript? Redaction runs before anything is stored (API keys, tokens, passwords, JWTs, private keys, connection strings, cookies, webhooks, phone numbers, emails). It is best-effort — treat the archive as sensitive data.

Why is my non-ASCII output path refused? SQLite on Windows cannot use it. The run aborts with exit 8; --allow-dir-fallback moves the archive to ~/.openclaw/sqlite-data and reports the change instead of doing it silently.

Issues & discussions

Sponsors

If you find this project helpful, consider supporting its development! See SPONSORS.md for donation options.

License

MIT License with a mandatory attribution requirement: using all or part of the source, including modified variants, is permitted — but Pondsi must always be credited as the original author. See LICENSE.

简体中文

这是什么？

InfinityContext 是一个可移植的 AI 智能体技能（标准 SKILL.md 包）：压缩上下文、把脱敏后的对话片段写入本地 SQLite/FTS5 归档，并可按需精确检索任意早期细节。它可运行在任何支持 Agent Skills 的宿主上——原生支持 DeepSeek Harness（dsh）（放进 .agents/skills/ 即被发现），同样支持 Claude Code、OpenClaw、Cursor、Dify、Ollama 及自研 Agent。

边界说明（说清楚）：本包只包含可移植的 Python 核心——读取会话轨迹、写入本地归档、提供检索。不联网、不调用 shell、不启动子进程、不常驻后台。它对文件能做的全部操作，都列在 SKILL.md 顶部的能力披露表里。

它做什么

步骤 脚本 说明
归档 scripts/session_to_sqlite.py 会话轨迹（JSONL）→ 脱敏片段写入本地 SQLite + FTS5 索引
检索 scripts/search.py 只读 FTS5 搜索（三元组；短中文词回退 LIKE）
清理 scripts/cleanup.py 归档文件按保留期清理，默认演练模式
保护 scripts/secure_fs.py 归档目录/数据库/WAL 旁文件强制 owner-only 权限

本包会执行的文件操作（完整披露见 SKILL.md 顶部表格）：

操作 工具 约束
把脱敏片段写入本地归档 session_to_sqlite.py owner-only 权限，Fail-Closed
永久删除过期归档文件 cleanup.py 归档标记 + 完整文件名白名单 + 不递归 + --apply --confirm-destructive
就地改写文件（脱敏） session_to_sqlite.py --redact-file 必须 --allow-dir；符号链接解析后的路径必须仍在范围内
路径非 ASCII 时移动归档 session_to_sqlite.py 默认拒绝，需显式 --allow-dir-fallback

保留期策略（看一次就够）

设置 取值 说明
默认保留 30 天 每次运行都会在同一事务内清理超过 30 天的片段
可配置范围 1..3650 天 --retention-days N
不限时间 必须显式开启 --retention-days 0 会被拒绝，除非同时给出 --allow-unbounded-retention
手动清理 --purge-only --output-dir <目录> 对既有归档执行同一策略，不摄入新数据
彻底关闭 INFINITY_CONTEXT_NO_ARCHIVE=1 归档器不写任何文件，返回 status: disabled

保留期是代码强制而非文档承诺：每次运行都会删除 session_chunks 及其 FTS 镜像中的过期片段，并在结果中返回 purged_chunks。

快速上手（5 分钟）

前置条件：Python 3.9+。不需要 Node.js、不需要构建、不需要联网。

# 1. 安装（注册表，已扫描产物）
clawhub install infinitycontext --workdir ~/.agents --dir skills   # dsh / Claude Code
clawhub install infinitycontext --workdir ~/.openclaw --dir skills  # OpenClaw

# 2. 归档一份会话轨迹
python3 scripts/session_to_sqlite.py \
  --session-key my-session \
  --session-file ~/path/to/events.jsonl \
  --output-dir ~/.infinity-context/archive

# 3. 以后精确检索（只读）
python3 scripts/search.py --db ~/.infinity-context/archive/my-session-*.db --query "部署令牌"

# 4. 清理归档文件（先演练）
python3 scripts/cleanup.py --archive-dir ~/.infinity-context/archive --dry-run

源码安装——固定已审计 tag 并逐文件校验：

git clone https://github.com/Pondsi/infinitycontext.git
cd infinitycontext
git checkout --detach v1.8.9
grep -q '^version: "1.8.9"' SKILL.md || { echo "tag/version mismatch - stop"; exit 1; }
sha256sum -c checksums.txt          # macOS：shasum -a 256 -c checksums.txt

随后按 references/architecture.md 列出的文件清单逐文件放入技能目录——清单是显式的，不需要通配符，也不需要 cp -r。

安装到哪

宿主 位置 说明
DeepSeek Harness（dsh） ~/.agents/skills/infinity-context/ 或 <项目>/.agents/skills/infinity-context/ 技能是一层深的目录包（<root>/<dir>/SKILL.md）；嵌套 **/SKILL.md 故意不被发现
OpenClaw <工作区>/skills/（优先级最高）或 ~/.openclaw/skills/ 自动发现
Claude Code ~/.claude/skills/infinity-context/ 同样的目录结构
Cursor / Dify / Ollama / 自研 让 Agent 指向本目录 纯 Python 标准库

系统要求

  • 任意支持标准 SKILL.md 的宿主：dsh、Claude Code、OpenClaw、Cursor、Dify、Ollama 或自研 Agent。
  • 只需 Python 3.9+。核心不含任何 JavaScript、PowerShell，不启动子进程，也没有网络代码。
  • 可选的 OpenClaw/Windows 自动化（钩子 + 看门狗）位于仓库 openclaw/ 目录，不属于本包。

常见问题

数据存在哪？ 在你传入的 --output-dir（默认 ~/.infinity-context/archive）：一个 SQLite 数据库及其 -wal/-shm 旁文件，owner-only 权限。

怎么彻底停止归档？ 设置 INFINITY_CONTEXT_NO_ARCHIVE=1，归档器不写任何文件并返回 status: disabled。

怎么只保留更少历史？ --retention-days 7，下次运行即删除超期片段；--purge-only --output-dir <目录> 可立即执行且不摄入新数据。

数据会外传吗？ 不会。包内没有任何网络代码。

轨迹里的密钥怎么办？ 脱敏在写入之前完成（API Key、Token、密码、JWT、私钥、连接串、Cookie、Webhook、手机号、邮箱），但它是尽力而为——请把归档当作敏感数据。

为什么我的非 ASCII 路径被拒绝？ Windows 上的 SQLite 无法使用该路径。运行会以退出码 8 中止；--allow-dir-fallback 会把归档改到 ~/.openclaw/sqlite-data，并如实上报，绝不静默改道。

问题与讨论

安全与隐私

  • 纯本地：无网络请求、无遥测、无云同步。
  • 无 shell、无子进程：核心脚本从不启动其它程序；Windows ACL 使用进程内 Win32 安全 API。
  • 归档仅本人可读，Fail-Closed：POSIX 0700/0600；Windows 受保护 DACL；加固后回读校验；无法强制 owner-only 时中止并销毁半成品（退出码 3）。
  • 脱敏 Fail-Closed：规则启动期校验并预编译，任何失败都在建库之前中止；session_key 在生成任何路径/文件名之前先净化。
  • 保留期有界：默认 30 天，范围 1..3650，无限期保留需显式 --allow-unbounded-retention。
  • 清理需二次确认：cleanup.py 只清理带 .infinity-context-archive 标记的目录，必须同时给出 --apply --confirm-destructive，不递归，通用 .json/.tmp/.bak 永不删除。

许可证

MIT 许可证（附强制署名条款）：允许使用全部或部分源码（含修改后的变体），但必须标注 Pondsi 的署名。详见 LICENSE。

繁體中文

這是什麼？

InfinityContext 是可移植的 AI Agent Skill（標準 SKILL.md）：壓縮上下文、把去識別化後的對話片段寫入本機 SQLite/FTS5 封存，並可精確檢索早期細節。支援 dsh、Claude Code、OpenClaw、Cursor、Dify、Ollama 與自訂 Agent。

安裝與使用

clawhub install infinitycontext --workdir ~/.agents --dir skills

完整步驟見 English 與 简体中文。

重點

  • 只需 Python 3.9+：純標準庫，不聯網、無 shell、不啟動子程序。
  • 保留期有界：預設 30 天（--retention-days 1..3650）；不限時間需顯式 --allow-unbounded-retention；INFINITY_CONTEXT_NO_ARCHIVE=1 可完全關閉封存。
  • 僅本機可讀：封存目錄 0700／檔案 0600（Windows 為受保護 DACL），Fail-Closed。
  • 清理需二次確認：cleanup.py 必須 --apply --confirm-destructive，且只清理帶標記的目錄。

授權

MIT（附強制署名條款）：可自由使用（含修改後的變體），但必須標註 Pondsi 的署名。

日本語

これは何？

InfinityContext は移植可能な AI エージェントスキル（標準 SKILL.md）です。コンテキストを圧縮し、秘匿化済みの会話断片をローカル SQLite/FTS5 アーカイブへ保存し、過去の詳細を正確に検索できます。dsh、Claude Code、OpenClaw、Cursor、Dify、Ollama、独自エージェントに対応。

インストールと使い方

clawhub install infinitycontext --workdir ~/.agents --dir skills

完全な手順は English と 简体中文 を参照してください。

要点

  • Python 3.9+ のみ：標準ライブラリのみ、ネットワークなし・shell なし・子プロセスなし。
  • 保持期間は有界：既定 30 日（--retention-days 1..3650）。無期限は明示的な --allow-unbounded-retention が必要。INFINITY_CONTEXT_NO_ARCHIVE=1 でアーカイブを完全に停止。
  • 所有者のみ読み取り可：ディレクトリ 0700／ファイル 0600（Windows は保護 DACL）、Fail-Closed。
  • 削除は二重確認：cleanup.py は --apply --confirm-destructive が必須で、マーカーのあるディレクトリだけを対象にします。

ライセンス

MIT（強制署名条項付き）：全体または一部（改変版を含む）の利用は自由ですが、Pondsi のクレジット表記が必須です。

한국어

무엇인가요?

InfinityContext는 이식 가능한 AI 에이전트 스킬(표준 SKILL.md)입니다. 컨텍스트를 압축하고 비식별화된 대화 조각을 로컬 SQLite/FTS5 아카이브에 저장하며, 과거의 세부 내용을 정확히 검색합니다. dsh, Claude Code, OpenClaw, Cursor, Dify, Ollama 및 사용자 정의 에이전트를 지원합니다.

설치와 사용

clawhub install infinitycontext --workdir ~/.agents --dir skills

전체 절차는 English와 简体中文를 참고하세요.

핵심

  • Python 3.9+만 필요: 표준 라이브러리만 사용, 네트워크 없음·shell 없음·하위 프로세스 없음.
  • 보존 기간은 유한: 기본 30일(--retention-days 1..3650). 무제한은 명시적 --allow-unbounded-retention이 필요하며 INFINITY_CONTEXT_NO_ARCHIVE=1로 완전히 끌 수 있습니다.
  • 소유자만 읽기 가능: 디렉터리 0700/파일 0600(Windows는 보호된 DACL), Fail-Closed.
  • 삭제는 이중 확인: cleanup.py는 --apply --confirm-destructive가 필요하며 마커가 있는 디렉터리만 정리합니다.

라이선스

MIT(필수 저작자 표시 조항): 전체 또는 일부(수정본 포함) 사용은 자유이지만 Pondsi를 반드시 명시해야 합니다.

Español

¿Qué es esto?

InfinityContext es una skill de agente de IA portátil (un paquete SKILL.md estándar): comprime el contexto, archiva fragmentos de conversación anonimizados en un almacén local SQLite/FTS5 y recupera detalles exactos de turnos anteriores. Funciona en dsh, Claude Code, OpenClaw, Cursor, Dify, Ollama y agentes propios.

Instalación y uso

clawhub install infinitycontext --workdir ~/.agents --dir skills

Los pasos completos están en English y 简体中文.

Puntos clave

  • Solo Python 3.9+: biblioteca estándar, sin red, sin shell, sin subprocesos.
  • Retención acotada: 30 días por defecto (--retention-days 1..3650); sin límite solo con --allow-unbounded-retention; INFINITY_CONTEXT_NO_ARCHIVE=1 detiene el archivado.
  • Solo el propietario puede leer: directorio 0700 / archivos 0600 (DACL protegida en Windows), fail-closed.
  • Borrado con doble confirmación: cleanup.py exige --apply --confirm-destructive y solo actúa sobre directorios con marcador.

Licencia

MIT (con cláusula de atribución obligatoria): se permite usar todo o parte, incluso variantes modificadas, pero Pondsi debe figurar siempre como autor.

Português

O que é?

InfinityContext é uma skill de agente de IA portátil (um pacote SKILL.md padrão): comprime o contexto, arquiva trechos de conversa anonimizados em um armazenamento local SQLite/FTS5 e recupera detalhes exatos de turnos anteriores. Funciona em dsh, Claude Code, OpenClaw, Cursor, Dify, Ollama e agentes próprios.

Instalação e uso

clawhub install infinitycontext --workdir ~/.agents --dir skills

O passo a passo completo está em English e 简体中文.

Pontos-chave

  • Apenas Python 3.9+: biblioteca padrão, sem rede, sem shell, sem subprocessos.
  • Retenção limitada: 30 dias por padrão (--retention-days 1..3650); sem limite apenas com --allow-unbounded-retention; INFINITY_CONTEXT_NO_ARCHIVE=1 desliga o arquivamento.
  • Somente o proprietário lê: diretório 0700 / arquivos 0600 (DACL protegida no Windows), fail-closed.
  • Exclusão com dupla confirmação: cleanup.py exige --apply --confirm-destructive e só age em diretórios com marcador.

Licença

MIT (com cláusula de atribuição obrigatória): é permitido usar tudo ou parte, inclusive variantes modificadas, mas Pondsi deve ser sempre creditado.

Français

Qu'est-ce que c'est ?

InfinityContext est une skill d'agent IA portable (un paquet SKILL.md standard) : elle compresse le contexte, archive des fragments de conversation anonymisés dans un stockage local SQLite/FTS5 et retrouve des détails exacts de tours antérieurs. Compatible dsh, Claude Code, OpenClaw, Cursor, Dify, Ollama et agents maison.

Installation et usage

clawhub install infinitycontext --workdir ~/.agents --dir skills

Les étapes complètes sont dans English et 简体中文.

Points clés

  • Python 3.9+ uniquement : bibliothèque standard, pas de réseau, pas de shell, pas de sous-processus.
  • Rétention bornée : 30 jours par défaut (--retention-days 1..3650) ; illimitée uniquement avec --allow-unbounded-retention ; INFINITY_CONTEXT_NO_ARCHIVE=1 désactive l'archivage.
  • Lecture par le propriétaire seul : répertoire 0700 / fichiers 0600 (DACL protégée sous Windows), fail-closed.
  • Suppression à double confirmation : cleanup.py exige --apply --confirm-destructive et n'agit que sur un répertoire portant le marqueur.

Licence

MIT (avec clause d'attribution obligatoire) : l'usage total ou partiel, variantes modifiées comprises, est autorisé, mais Pondsi doit toujours être crédité.

Deutsch

Was ist das?

InfinityContext ist eine portable KI-Agent-Skill (ein Standard-SKILL.md-Paket): Sie komprimiert den Kontext, archiviert entidentifizierte Gesprächsausschnitte in einem lokalen SQLite/FTS5-Speicher und ruft exakte Details früherer Turns ab. Läuft auf dsh, Claude Code, OpenClaw, Cursor, Dify, Ollama und eigenen Agenten.

Installation und Nutzung

clawhub install infinitycontext --workdir ~/.agents --dir skills

Die vollständigen Schritte stehen in English und 简体中文.

Kernpunkte

  • Nur Python 3.9+: Standardbibliothek, kein Netzwerk, keine Shell, keine Subprozesse.
  • Begrenzte Aufbewahrung: standardmäßig 30 Tage (--retention-days 1..3650); unbegrenzt nur mit --allow-unbounded-retention; INFINITY_CONTEXT_NO_ARCHIVE=1 schaltet die Archivierung ab.
  • Nur der Eigentümer liest: Verzeichnis 0700 / Dateien 0600 (unter Windows geschützte DACL), fail-closed.
  • Löschen nur mit doppelter Bestätigung: cleanup.py verlangt --apply --confirm-destructive und arbeitet nur in einem Verzeichnis mit Marker.

Lizenz

MIT (mit verpflichtender Namensnennung): Nutzung ganz oder teilweise, auch modifiziert, ist erlaubt — Pondsi muss jedoch immer genannt werden.

Русский

Что это?

InfinityContext — переносимый навык ИИ-агента (стандартный пакет SKILL.md): сжимает контекст, сохраняет обезличенные фрагменты диалога в локальное хранилище SQLite/FTS5 и точно находит детали прошлых ходов. Работает в dsh, Claude Code, OpenClaw, Cursor, Dify, Ollama и собственных агентах.

Установка и использование

clawhub install infinitycontext --workdir ~/.agents --dir skills

Полные шаги — в разделах English и 简体中文.

Ключевое

  • Только Python 3.9+: стандартная библиотека, без сети, без shell, без подпроцессов.
  • Ограниченное хранение: по умолчанию 30 дней (--retention-days 1..3650); без ограничения — только с явным --allow-unbounded-retention; INFINITY_CONTEXT_NO_ARCHIVE=1 полностью отключает архивирование.
  • Чтение только владельцем: каталог 0700 / файлы 0600 (в Windows — защищённый DACL), fail-closed.
  • Удаление с двойным подтверждением: cleanup.py требует --apply --confirm-destructive и работает только в каталоге с маркером.

Лицензия

MIT (с обязательным указанием авторства): использование целиком или частично, включая изменённые варианты, разрешено, но Pondsi должен быть указан всегда.

Security / 安全模型

Fail-closed by default — nothing is written when a safety check cannot be enforced.

  • Local only: no network code, no telemetry, no MCP, no cloud sync anywhere in the package.
  • No shell, no subprocesses: the four Python scripts never start another program; Windows ACL hardening uses in-process Win32 security APIs.
  • Owner-only archive, fail-closed: the archive directory is forced to 0700 and files to 0600 (POSIX) or a protected DACL (Windows); the result is re-read to prove it took effect. If owner-only access cannot be enforced, archiving aborts and the half-written database is destroyed (exit 3). --allow-insecure-storage is the only opt-out and is reported as insecure_storage: true.
  • Fail-closed redaction: rules are validated and precompiled at startup; a malformed rule file aborts before any database is created. session_key is sanitised before it is used for any path or filename.
  • Bounded retention: 30 days by default, --retention-days 1..3650; unlimited retention requires the explicit --allow-unbounded-retention flag; INFINITY_CONTEXT_NO_ARCHIVE=1 disables archiving entirely.
  • Deny-by-default filesystem rules: cleanup.py refuses a directory without the .infinity-context-archive marker, deletes only full filenames matching InfinityContext artifacts, never recurses, re-checks every candidate with lstat, and does nothing without --apply --confirm-destructive.

默认拒绝 — 任何安全检查无法强制执行时，直接中止而不是降级。

  • 纯本地：包内没有任何网络代码、遥测、MCP 或云同步。
  • 无 shell、无子进程：四个 Python 脚本从不启动其它程序；Windows ACL 使用进程内 Win32 安全 API。
  • 归档仅本人可读，Fail-Closed：POSIX 0700/0600，Windows 受保护 DACL，并回读校验；无法强制时中止并销毁半成品（退出码 3）。唯一例外是 --allow-insecure-storage，且结果中会标记 insecure_storage: true。
  • 脱敏 Fail-Closed：规则启动期校验并预编译，规则文件损坏会在建库前中止；session_key 在生成任何路径/文件名之前先净化。
  • 保留期有界：默认 30 天，--retention-days 1..3650；无限期保留需显式 --allow-unbounded-retention；INFINITY_CONTEXT_NO_ARCHIVE=1 可彻底关闭归档。
  • 默认拒绝的文件系统规则：cleanup.py 拒绝没有 .infinity-context-archive 标记的目录，只删除完整文件名匹配本应用产物的文件，不递归，删除前用 lstat 复核，且必须同时给出 --apply --confirm-destructive。

隐私声明 / Privacy Statement

本 Skill 会将对话上下文保存在本地纯内网的 SQLite 中以供检索，系统已内置正则脱敏机制屏蔽常见 API 密钥，且不依赖任何云端同步。所有数据仅存储于本机，不会外传。

This skill stores conversation context in a local-only SQLite database for retrieval. Built-in regex redaction masks common API keys. No cloud sync is used; all data stays on the local machine.

Pondsi (+MiMo-v2.5/v2.5pro+deepseek-v4-flash/pro+deepseek-v4.1-flash-expires-on-0910+GLM5.3-flash+Gemini3.1-pro+Qwen3.8-27b+Gemini3.8-flash) — automatically committed by Openclaw

About

Open-Source Context Compression & Memory Optimization for OpenClaw

Topics

ai-agentcontext-compressionfts5javascriptllmmemory-optimizationmulti-agentpowershellpythonsqlite

Resources

Readme
License

Security policy

Security policy
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages