InfinityContext

Your agent never forgets. Compress the context window without losing a single detail — and know exactly when and how to bring any of it back.

Long sessions eventually 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. When something from hundreds of turns ago matters again, the agent searches the archive and returns the exact message — offline, on your machine.

Before / after a context compaction

Scope, stated plainly. Details stay retrievable inside the retention window — 30 days by default, configurable 1..3650, or unbounded only with an explicit flag. No network, no shell, no subprocesses.

⚠️ Security & Privacy Disclosure (Intended Behavior)

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 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
Enforce a bounded retention window (default 30 days)	session_to_sqlite.py	--retention-days 1..3650; keeping chunks forever needs --allow-unbounded-retention
Disable archiving entirely	session_to_sqlite.py	env INFINITY_CONTEXT_NO_ARCHIVE=1 — the script then writes nothing

Nothing is sent anywhere: no network, no telemetry, no cloud sync. Redaction is best-effort; the archive still holds a detailed record of your sessions, so keep it out of synced or shared folders and run cleanup deliberately. Retention is bounded by default: chunks older than 30 days are purged on every run, and --purge-only applies the same policy on demand.

Installing this skill means accepting these local persistence and file-mutation capabilities.

Works out of the box on DeepSeek Harness (dsh) and OpenClaw, and on any host that loads a standard SKILL.md.

Install

1. Registry (scanned artifact — no git, no build step)

# dsh / Claude Code: lands in ~/.agents/skills/infinity-context/
clawhub install infinitycontext --workdir ~/.agents --dir skills

# OpenClaw: managed skills (~/.openclaw/skills) or workspace skills (higher precedence)
clawhub install infinitycontext --workdir ~/.openclaw --dir skills
clawhub install infinitycontext --workdir <workspace> --dir skills

2. From source — pin the reviewed release tag, verify every byte

git clone https://github.com/Pondsi/infinitycontext.git
cd infinitycontext
git checkout --detach v1.8.7
grep -q '^version: "1.8.7"' SKILL.md || { echo "tag/version mismatch - stop"; exit 1; }
sha256sum -c checksums.txt                # macOS: shasum -a 256 -c checksums.txt
# compare the output with the hashes published in the GitHub release notes

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

The pinned tag must equal the version in this file's frontmatter. The grep guard above stops the install when it does not, so a source install can never silently produce an older build than the reviewed artifact.

Quick start on DeepSeek Harness (dsh)

dsh uses the standard SKILL.md contract. Two rules matter:

• frontmatter name and description are required, and name must be kebab-case
• a skill is a directory bundle one level deep (<root>/<dir>/SKILL.md) or a flat <name>.md; nested **/SKILL.md files are deliberately not discovered
• the folder name does not have to match name — dsh identifies the skill by the frontmatter name, so the registry bundle infinitycontext/ works as-is

Use the user-level root ~/.agents/skills/infinity-context/ (rank 500, shared with Claude Code and other agents) or the project-level root <project>/.agents/skills/infinity-context/ (rank 200, wins over the user-level copy). Restart dsh, type /, and the skill appears under Skills. Nothing else to install: a skill takes effect the moment its folder sits in a scan root.

Then drive it from the agent's shell tool:

# archive a transcript (JSONL) into the local SQLite/FTS5 store
python3 scripts/session_to_sqlite.py --session-key <key> --session-file <events.jsonl> --output-dir ~/.infinity-context/archive

# retrieve a detail from months ago
python3 scripts/search.py --query "deployment token"

# prune old archives (dry-run by default)
python3 scripts/cleanup.py --dry-run

Quick start on OpenClaw

The scripts above are then available unchanged; OpenClaw discovers the skill from its skills directory and calls them through its shell tool. Host-specific automation is optional and lives outside this package (see the repository).

Host compatibility

The core is host-agnostic: four Python scripts, no network, no shell, no subprocesses, no Windows-only dependency.

What it does

1. Archive (session_to_sqlite.py) — turns a session transcript into a local SQLite database with an FTS5 index, so a compressed session can still be searched down to the message.
2. Retrieve (search.py) — FTS5 trigram search with a LIKE fallback for short CJK queries; reads only, never writes.
3. Prune (cleanup.py) — retention-based cleanup with canonical path anchoring and VACUUM; dry-run by default.
4. Protect (secure_fs.py) — owner-only permissions for the archive directory, the database and its WAL sidecars.

Compression itself is a prompt-level discipline: keep keepRecentTokens small enough that a deep reply still fits, and let the archive carry the details instead of the context window.

First-run consent (agent behaviour)

Before archiving a session for the first time in a given environment, the agent must tell the user that the conversation will be stored locally in a searchable archive, and obtain explicit confirmation. Do not archive silently. When a user asks to stop keeping history, set INFINITY_CONTEXT_NO_ARCHIVE=1 (the archiver then writes nothing) and run cleanup.py --apply --confirm-destructive — or delete the archive directory. Retention is bounded by default (30 days), so history stays inside a known window.

Security & privacy

• Local only — no network calls, no telemetry, no MCP, no cloud sync.
• No shell, no subprocesses — the core never starts another program. Windows ACL hardening uses in-process Win32 security API calls, not a helper executable.
• Owner-only archive, fail-closed — the archive directory is forced to 0700 and files to 0600 on POSIX; on Windows the DACL is replaced by a protected DACL granting only the current user and LOCAL SYSTEM. Every result is re-read to prove the mode took effect. New database files are created atomically with O_CREAT | O_EXCL | O_NOFOLLOW and 0600, so no file ever exists with wider permissions. A symbolic link on the target path is refused. A pre-existing directory owned by another account is refused. If owner-only access cannot be enforced, archiving aborts and the half-written database is destroyed (status: error, exit 3) instead of storing readable data; --allow-insecure-storage is the only way to opt out, and the JSON result then reports insecure_storage: true.
• Fail-closed redaction — before any text is stored, a regex redactor masks API keys, tokens, passwords, JWTs, private keys, connection strings, cookies, webhooks, phone numbers and emails. Rules are validated and precompiled at startup: a malformed redact_rules.json, a wrong field type or an uncompilable regex aborts the run before any database is created, and a failure while applying a rule aborts instead of skipping it. session_key is sanitised before it is used for any path or filename — a value outside the safe identifier format (or one that itself looks sensitive) becomes an opaque hash, so it never reaches a filename, the table or the FTS index. The transcript is redacted entirely in memory, then written in a single transaction; on failure the transaction rolls back and only a database created by that same run is removed — an existing archive being appended to is never deleted. High-entropy candidates are excluded from the keyword index. In-place redaction (--redact-file) additionally requires --allow-dir and resolves every symbolic link before comparing paths: the lexical path, the resolved path and the resolved allowed directory must all agree, so a symlinked ancestor inside the allowed directory cannot redirect the write elsewhere. A non-ASCII output path is refused by default (exit 8); only an explicit --allow-dir-fallback moves the archive to the ASCII fallback directory, and the run then prints a warning and reports archive_dir_fallback: true together with requested_dir and archive_dir — the location is never changed silently.
• Data minimisation — MAX_ARCHIVE_LENGTH truncates oversized content (head + tail kept) before storage. Ingestion itself is bounded before parsing: at most --max-session-bytes (64 MiB) is read from the head of the transcript, a line longer than --max-line-bytes (1 MiB) is discarded before JSON or any regex sees it, and ingestion stops at --max-messages (200000) or --max-total-chars (64 MiB). The result reports ingest.truncated and ingest.truncated_reason, so a bounded archive is never presented as a complete one. In-place redaction refuses a file larger than 64 MiB before reading it.
• Bounded retention, default 30 days — archiving is not unbounded. Every run purges chunks older than --retention-days (default 30, range 1..3650) from session_chunks and its FTS mirror inside the same transaction, and reports the count as purged_chunks; --purge-only --output-dir <dir> applies the same policy to existing archives without ingesting, and only to a directory that carries the archive marker. Keeping chunks forever requires the explicit --allow-unbounded-retention flag. Set INFINITY_CONTEXT_NO_ARCHIVE=1 to disable archiving entirely — the script writes no file and reports status: disabled.
• Archive files are created exclusively and identified before they are written — a new archive is reserved with O_CREAT | O_EXCL | O_NOFOLLOW and 0600, so a name collision aborts (exit 10) instead of reusing an existing file. --purge-only only considers the complete artifact filename pattern and requires a read-only identity match (archive_metadata app / format version / archive id, plus the expected session_chunks columns) before it re-opens a database read-write; skipped databases are reported, never modified.
• Append is identity-checked, not prefix-matched — --append only accepts a file whose complete name is {safe_key}-YYYYMMDD-HHMMSS.db, opens it read-only first and requires the exact session_key in archive_metadata (older archives fall back to a row check), then upgrades the archive with that metadata row. A foreign database, a symlink, a path outside --output-dir, or more than one candidate all abort with exit 9 before anything is written; --db-path picks one explicitly.
• Deny-by-default filesystem rules — cleanup.py refuses any directory that lacks the owner-only .infinity-context-archive marker, refuses protected directories (filesystem root, home, common user folders), only deletes files whose full name matches an InfinityContext artifact pattern, never recurses into subdirectories, re-checks each candidate with lstat immediately before deletion, validates the session_chunks/chunk_fts schema in read-only mode before any VACUUM, and does nothing unless both --apply and --confirm-destructive are given.

Data sensitivity notice

Conversation archives contain session history and are treated as sensitive data. Redaction is best-effort: after redaction the archive still holds a detailed record of your sessions. Do not share archive files, sync them to cloud storage, or widen the permissions of the archive directory without understanding the consequences.

Supply chain

This package is self-contained: SKILL.md, README.md, 说明.md, CHANGELOG.md, SPONSORS.md, LICENSE, checksums.txt, scripts/ (Python only), references/, sponsors/. It contains no JavaScript, no PowerShell and no code fetched at install time — the audited artifact is exactly what runs. checksums.txt lists the SHA-256 of every published file except itself.

Any host-specific automation is deliberately out of scope for this package. Anything of that kind lives in the repository outside the published artifact and carries its own documentation, pinned revision and checksums.

Want the complete project? The GitHub repository (Pondsi/infinitycontext) is the full, unabridged version — the identical core plus the optional host integration (openclaw/, with its own README and checksums), so nothing is missing when you need it. The registry package here is deliberately slimmed to the auditable core.

Configuration

Architecture

session transcript (JSONL)
        │
        ├─ bound  ────────────────► size / line / count / char caps, applied while reading
        ├─ redact  ────────────────► fail-closed: nothing is stored if this step fails
        ├─ truncate (MAX_ARCHIVE_LENGTH)
        ▼
  session_chunks  ──trigger──►  chunk_fts (FTS5 trigram)
        │
        └─ search.py  ──►  exact detail from any past turn

More languages

Localised summaries (繁體中文, 日本語, 한국어, Español, Português, Français, Deutsch, Русский) are in references/languages.md.

License

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

无限上下文压缩与记忆归档

让智能体永不健忘。 压缩上下文窗口，却一个细节都不丢——并且知道什么时候、用什么方法，把哪个细节找回来。

长会话迟早会撞上上下文上限。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，智能体检索归档，取回那一句原话——离线、在本机。

上下文压缩前后对比

范围说明（如实）：细节在保留窗口内可检索——默认 30 天，可配 1..3650，无界需显式开关。不联网、无 shell、无子进程。

开箱即用：DeepSeek Harness（dsh）与 OpenClaw；任何加载标准 SKILL.md 的宿主亦可。

安装

# 方式一：注册表（已扫描产物，无需 git、无需构建）
clawhub install infinitycontext --workdir ~/.agents --dir skills    # dsh / Claude Code
clawhub install infinitycontext --workdir ~/.openclaw --dir skills  # OpenClaw 托管技能

方式二（源码安装：固定已审计 tag + 逐文件校验）见上方 Install； 需要复制的文件清单见 references/architecture.md。

必须逐文件显式复制，禁止 cp -r：被审计的包不得混入未审计文件。

在 dsh 上快速开始

• frontmatter 的 name 与 description 必填，且 name 必须是 kebab-case
• 技能是一层深的目录包（<root>/<dir>/SKILL.md）或平铺文件 <name>.md；嵌套的 **/SKILL.md 故意不被发现
• 文件夹名不必与 name 相同——dsh 用 frontmatter 的 name 作为标识（注册表安装出的目录是 infinitycontext/，同样可用）

用户级 ~/.agents/skills/infinity-context/（rank 500，与 Claude Code 共享）或项目级 <project>/.agents/skills/infinity-context/（rank 200，优先）。重启 dsh，输入 /， 技能即出现在 Skills 分组——文件夹进入扫描根即生效，无需其它步骤。

在 OpenClaw 上快速开始

clawhub install infinitycontext 后，上述脚本即可直接调用；OpenClaw 从技能目录发现本技能 并通过 shell 工具执行。宿主专有的自动化属于可选集成，不在本包内（见仓库）。

四个脚本

1. 归档 session_to_sqlite.py——会话轨迹转本地 SQLite + FTS5 索引
2. 检索 search.py——FTS5 三元组搜索，短中文词回退 LIKE；只读
3. 清理 cleanup.py——按保留期清理并 VACUUM；默认演练模式
4. 保护 secure_fs.py——归档目录/数据库/WAL 旁文件强制 owner-only 权限

安全与隐私

• 纯本地：不联网、无遥测、无 MCP、不上传
• 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，不调用外部工具
• 归档仅本人可读，且 Fail-Closed：POSIX 目录 0700、文件 0600；Windows 用受保护 DACL 仅授权当前用户与 LOCAL SYSTEM；加固后回读校验是否真的生效；新数据库文件以 O_CREAT|O_EXCL|O_NOFOLLOW + 0600 原子创建；路径上出现符号链接即拒绝；预存目录若属于其它账号则拒绝使用；无法强制 owner-only 时中止归档并销毁半成品（status: error，退出码 3），绝不留下可读的明文；仅 --allow-insecure-storage 可显式降级，且结果中会标记 insecure_storage: true
• Fail-Closed 脱敏：入库前屏蔽 API Key / Token / 密码 / JWT / 私钥 / 连接串 / Cookie / Webhook / 手机号 / 邮箱；规则启动期校验并预编译——redact_rules.json 损坏、字段类型错误或正则无法编译都会在建库之前中止，应用规则时出错也中止而非跳过；session_key 在生成任何路径/文件名之前先净化（不符合安全字符集或本身疑似敏感→不透明哈希），绝不进入文件名、表或 FTS；轨迹先在内存中全量脱敏，再在单事务内写入，失败则回滚且只清理本轮新建的库，追加模式下的历史归档绝不被删除；高熵候选不进入关键词索引；原地脱敏（--redact-file）必须同时给出 --allow-dir，且在比较路径前解析全部符号链接：词法路径、真实路径、真实允许目录三者必须一致，因此允许目录内部的符号链接祖先无法把写入重定向到别处；若因非 ASCII 路径回退到备用目录，会打印告警并在结果中给出 archive_dir_fallback: true、requested_dir 与 archive_dir——绝不静默改道
• 保留期有界，默认 30 天：归档不会无限累积。每次摄入都会在同一事务内删除超过 --retention-days（默认 30，范围 1..3650）的旧片段及其 FTS 镜像，并在结果中给出 purged_chunks；--purge-only --output-dir <目录> 可对既有归档执行同一策略（仅限带归档标记的目录，不摄入新数据）。要无限期保留必须显式 --allow-unbounded-retention。设置 INFINITY_CONTEXT_NO_ARCHIVE=1 可彻底关闭归档——脚本不写任何文件并返回 status: disabled。
• 数据最小化：MAX_ARCHIVE_LENGTH 掐头去尾截断
• 默认拒绝的文件系统规则：cleanup.py 只删归档目录内、白名单扩展名、非符号链接的文件，且必须显式 --apply

数据敏感性声明

归档包含会话历史，属于敏感数据。脱敏是尽力而为：脱敏后仍保留会话的详细记录。 请勿分享归档文件、同步到云存储，或在未理解后果的情况下放宽归档目录权限。

供应链

本包自包含：SKILL.md、README.md、说明.md、CHANGELOG.md、SPONSORS.md、 LICENSE、checksums.txt、scripts/（仅 Python）、references/、sponsors/。 不含任何 JavaScript / PowerShell，也不在安装时拉取外部代码——被审计的产物就是 实际运行的东西。checksums.txt 列出除自身外每个发布文件的 SHA-256。

宿主专有的自动化不属于本包范围，只存在于仓库中、位于发布产物之外， 并自带文档、固定版本号与校验和。

想要完整项目？ GitHub 仓库 （Pondsi/infinitycontext）是完整、不精简的版本： 核心完全一致，外加可选的宿主集成（openclaw/，自带 README 与校验和），需要时功能一个不少。 你在这里看到的注册表包刻意精简为可审计核心。

其它语言

繁體中文、日本語、한국어、Español、Português、Français、Deutsch、Русский 见 references/languages.md。

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。

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