Time Library
Time Library / 时间图书馆 is the product name.
Time Library is easiest to understand as a five-step local workflow:
capture -> recall -> answer from evidence -> install agent rule -> health
It is not a hosted chat app and not a summary-to-vector API. It keeps original records on your machine, returns source refs before raw excerpts, and gives local agents a standing rule for when to check memory before they answer or act.
Why Time Library (not just another memory store)
Three things most memory tools don't do:
- Source-provenance, not paraphrase. Every remembered card traces back to the byte offset of the original words. Summaries are navigation; the raw record is the authority. Not "I think I remember" — "here's the original, at this line."
- It keeps time, not just text. Ordered timestamps and the raw → daily → digest sediment preserve the foundation for historical views. A public as-of / time-travel query endpoint does not ship yet.
- Many agents, one low-noise pool — no human relay. On one machine, Codex / Claude / OpenClaw / Hermes read a shared, read-only, traceable project memory. It hands out a catalog (a map), not a dump into your window, so recall stays low-contamination — and you stop being the middleman who repeats context between tools.
Why not just use Cognee (or Mem0)? The design priority is different: Time Library keeps the original source and timestamp, shares one local pool across agents, and carries delivery/use receipt fields. Those fields do not yet prove model-in-the-loop delivery on every platform. Provenance is not decoration: in the project's ablation, removing the two-layer / source path costs 13.31 points.
Core Workflow
- Capture source records: keep original conversations, tool output, source tool, device, and timeline before any summary. Summaries help navigation, but original records remain the source of truth.
- Recall with source refs: ask about old decisions, preferences, fixes, project boundaries, or next steps and get compact source refs, library ids, hit reasons, and optional bounded excerpts.
- Answer from evidence: when a model is configured, Time Library can ask it to answer only from supplied evidence, cite supporting refs, or return
UNKNOWNwhen evidence is insufficient. - Install an agent rule: add the Time Library skill/instruction or
time-libraryMCP tool so Codex, Claude, OpenClaw, Hermes, Cursor-style tools, and other local agents know when to call recall. - Check health before trust: use capability check, Record Doctor, and preflight doctor so install checks, daily recall, and troubleshooting do not blur together.
Advanced Capabilities
- Shared local context: use one local record base across Claude Desktop, Claude Code CLI, Codex, OpenClaw, Hermes, Cursor-style tools, and popular open-source agents.
- Preference and work-experience shelves: The preference shelf keeps preference and intent experience; The work-experience shelf keeps work experience, validation paths, gotchas, and repair order. Experience is not a skill library.
- Experience for every local agent: deliver source-backed work experience through skills, custom instructions, MCP, and
work_preflightso agents can check work history before acting. - Hermes skill evolution: compare Hermes skills with work-experience records in a read-only diff, then turn new skills or changed skills into reviewable adoption or upgrade candidates.
- Safe agent authority: memory is passive by default. Recall context cannot silently become a direct answer, and a direct answer cannot silently become a platform action.
- Local console: open a browser page to see tools detected on this machine, recent record health, safe capability checks, and where new raw records are stored.
- Local diagnostics: keep health checks, record checks, and troubleshooting reports separate from the daily recall path.
- Simple install options: use one shell command, PowerShell, or the double-click installers included in the release zip.
Reading Room and Whiteboard
The reading room is a shared workspace for several agents on one machine; the whiteboard inside it is the project progress board. Project state, in-flight tasks, and review boundaries live on one picture, so a new agent can see where a project actually stands instead of asking you.
Delivered and Used, Not Just Stored
Having a card in the library is not the same as an agent seeing it. Time Library can record surfacing receipts and borrow / return notes — useful, useless, or misleading — for the next experience candidate. The receipt path exists, but model-in-the-loop delivery is currently unproven on all seven tracked platform adapters (0/7); the store → use loop remains an acceptance target, not a finished claim.
Hermes: Autonomous Experience Upgrade
A controlled run-once proved the core Hermes path: new raw passed the gate, Hermes wrote its own skill, and Time Library produced a read-only, source-backed experience candidate. The OS runner and its hourly wake, 24-hour minimum interval, and one-run-per-day budget are implemented and source-tested. This release does not claim a separately observed unattended background fire-through, nor automatic adoption into production experience.
Quick Demo
After install, open the local console:
http://127.0.0.1:9850
Then run the safe first check:
{"query":"capability check","mode":"capability_check"}
A healthy first result says the connection is read-only, no real memory was recalled, and no raw excerpt was returned. After that, try a real question such as:
What did we decide last time about this project?
Time Library is designed to answer with source refs first, then expand to original evidence only when you ask.
Before asking an agent to change code, install, sync, or troubleshoot, ask it to check local context first. The expected behavior is simple: tell you whether the work looks already built, miswired, missing a diagnostic, or truly missing, then inspect the repo and tools before editing.
What It Remembers
AI tools forget the small things that make work smooth: your preferred wording, project boundaries, old mistakes, useful fixes, and where a task left off. Time Library keeps that trail on your own machine so a new agent window does not have to start from zero.
It is not a hosted chat app and not a summary vault. It keeps source records, source refs, corrections, and work experience together so memory can point back to the original words.
How Experience Evolves
Experience evolves, but it is not a black box. Time Library supports evidence-backed curation with validation and receipts. Experience moves like a library curation workflow:
raw record
-> experience candidate
-> review queue
-> source and acceptance-check validation
-> authorized adoption or rejection
-> rollback, supersede, or upgrade receipt
That means a useful repair path can become reusable work experience, while a bad or unsupported lesson can stay in review, move to errata, or be rolled back. The current system supports curated evolution; it does not claim fully autonomous self-training.
What You Get In Practice
- Shared local context for your AI tools: Claude Desktop, Claude Code CLI, Codex, OpenClaw, Hermes, Cursor-style tools, and fast-moving open-source agents can connect to the same local record base.
- Working methods that survive the next window: preferences stay available, while proven ways of working become reusable guidance.
- Preferences and experience stay distinct: the preference layer keeps preferences, corrections, and boundaries; the work-experience layer keeps repair paths, validation steps, and work methods.
- Experience can intervene across platforms: work experience is not private to one tool. Any local agent with a skill, custom instruction, or MCP entry can read the same experience candidates, gotchas, and acceptance checks before work.
- Fewer repeated fixes: before starting work, an agent can check whether you already built the feature, tested the path, or found the same wiring problem earlier.
- Receipts, not vibes: remembered items keep source trails so you can check why something came back.
- Original records stay original: summaries can help navigation, but they do not replace saved source text.
- Tool boundaries stay visible: Claude Desktop and Claude Code CLI are treated separately; relay or official-login records are not silently mixed.
- A safer first test: capability check proves the local read-only path works without recalling real memory or returning raw excerpts.
- Install once, then it finds your tools: Time Library discovers local AI tools, including popular open-source agents such as OpenCode, Goose, Aider, and OpenHands, connects usable local AI tool entries, and keeps new memory sources in a computer-first archive.
- Smarter recognition for new tools: if you have a model configured, or already have a local AI model setting Time Library can reuse, it can identify unfamiliar tools from metadata only; if not, local rules still work.
Paste This To Your Local Agent
If you use Codex, Claude Code CLI, OpenClaw, Hermes, or another local agent that can run terminal commands, paste this prompt into it:
You are installing Time Library for me on this machine.
Repository: https://github.com/strmforge/time-library
First check whether Time Library is already installed and running on this machine. Read `<TIME_LIBRARY_ROOT>/runtime/front_door_port` and probe the discovered front door and any local install marker you know how to inspect. If it is already installed and reachable, do not reinstall it. If it is not installed, install and start Time Library. In either case, connect this host only through capabilities the host actually reports; do not ask Time Library for a product-name adapter.
Then install Time Library as a standing memory rule for this agent, not just a one-time setup note. Before changing host configuration, inspect what this host supports. The host report is authoritative; product-name inference is only a hint. Use the documented, reversible configuration path owned by the host. If it supports MCP, register the MCP tool named time-library through the discovered front door at /mcp?startup_catalog=deferred. If it supports a skill or custom-instruction surface, add the Time Library skill/instruction. If it exposes a prompt/preflight hook, connect that supported hook to the local Time Library preflight path; the hook must stay quiet and never block prompts when the discovery file is unavailable. Record truthful source_system, client/window or session identity, skill_surface_status, config_write_authority, and current reading-area/project/series scope. Keep unsupported capabilities explicitly unsupported.
After setup, run only a safe capability check with time_library_recall:
{"query":"capability check","mode":"capability_check"}
Do not recall my real memory yet; do not recall my real memory until I ask.
Capability check is not connection proof. Later, only when one of my questions actually needs prior context and therefore authorizes real recall, call time_library_recall and retain a returned ZX-* library_id. In the same initialized MCP session, call time_library_reading_area with action self_report_connect, using the host-reported source_system, current canonical_window_id or session_id, truthful skill_surface_status and config_write_authority, one truthful reading_area/declared_project_ids/declared_series_ids scope, and proof_library_id set to that returned library_id. If any required fact or proof is missing, leave verification pending; never invent it. Then repeat the authorized recall through the verified connection. Handle a delivery challenge only after the host model actually received and used the cited refs. Never recall private memory merely to prove installation.
Keep this rule active from now on. When my question depends on prior context, call time_library_recall before answering and use source refs by default; ask for raw excerpts only when I explicitly need original evidence text. Triggers include previous decisions, corrections, project boundaries, forgotten context, install/test/release status, and short follow-ups in ongoing work such as "next step", "what else", or "then what". Use active recall by default: current window/session first, then same project/workspace, same workstream/task, then stable preferences/tool facts. Use raw-pool/global only when I explicitly ask for that wider view. If explicit window-only recall says the current window/session is not bound yet, say that plainly; do not claim there is no memory. If the skill is installed but time_library_recall is not available, tell me the MCP/tool connection is missing instead of guessing from memory.
Before coding, installing, syncing, release-prep, or remote troubleshooting, run a read-only pre-work check:
{"query":"<the work to do>","mode":"work_preflight"}
Use it to decide whether the work looks already built but forgotten, already built but miswired, missing diagnostics, or actually missing. Treat that result as a starting point; still inspect the repo, tests, tools, and docs before editing.
Chinese prompt:
你正在帮我在这台机器安装 Time Library。
仓库：https://github.com/strmforge/time-library
请先检查这台机器上的 Time Library 是否已经安装并在运行。先读取 `<TIME_LIBRARY_ROOT>/runtime/front_door_port` 并探测发现的门面，也检查你已知的本地安装标记。如果已经安装且可达，就不要重装；如果没装，再安装并启动 Time Library。无论哪种情况，都只按宿主真实自报的能力接入，不要让 Time Library 按产品名新增适配器。
然后把 Time Library 安装成这个 agent 的长期记忆规则，不只是一次性安装说明。修改宿主配置前，先检查并自报宿主支持哪些能力；宿主自报是权威，按产品名推断只能作提示。只使用宿主自己提供的、可回滚的正式配置路径。支持 MCP 就注册名为 time-library 的 MCP 工具，通过发现文件连接门面的 /mcp?startup_catalog=deferred；支持 skill 或自定义指令就添加 Time Library skill/指令；支持 prompt/preflight hook 就按宿主正式方式把它接到本机 Time Library preflight，且发现文件不可读时必须静默软失败，不能刷屏、不能卡 prompt。如实记录 source_system、当前窗口或 session 身份、skill_surface_status、config_write_authority，以及当前 reading-area/project/series 范围；不支持的能力明确保留为不支持。
安装完成后，只做一次安全能力检查：
{"query":"capability check","mode":"capability_check"}
先不要召回我的真实记忆。
capability check 不是接通证明。以后只有当我的真实问题确实依赖旧上下文、因而授权真实召回时，才调用 time_library_recall，并保留返回的 ZX-* library_id。在同一个已 initialize 的 MCP session 中调用 time_library_reading_area，action 使用 self_report_connect，如实填写宿主自报的 source_system、当前 canonical_window_id 或 session_id、skill_surface_status、config_write_authority、一个真实的 reading_area/declared_project_ids/declared_series_ids 范围，并把 proof_library_id 设为刚才返回的 library_id。缺任何必需事实或证明时都保持待验证，不得编造。随后通过已验证连接重放这次获授权的召回；只有宿主模型真正收到并使用了引用来源后，才处理 delivery challenge。不得为了证明安装成功而擅自召回私有记忆。
以后请持续遵守这条规则。凡是我的问题依赖旧上下文，请先调用 time_library_recall，默认结合 source_refs 回答；只有我明确需要原文证据时，才请求 raw_excerpt。触发词包括：之前、定论、纠错、边界、忘了、安装/测试/发布状态，以及持续项目里的“下一步/接下来呢/还有吗/然后呢”。默认使用 active 召回：当前窗口/session 优先，然后同项目/同工作区、同工作流/同任务、稳定偏好/工具事实。只有我明确要求更宽视图时，才使用 raw-pool/global。如果显式 window-only 召回提示当前窗口/session 还没绑定，请直接说明这个绑定缺口；不要说没有记忆。如果 skill 已安装但 time_library_recall 不可用，请告诉我 MCP/工具连接还没接上，不要凭印象猜。
写代码、安装、同步、发版准备或远端排障前，先做只读的开工前检查：
{"query":"<准备做的事>","mode":"work_preflight"}
用它先判断这件事更像已经做了但忘了、已经做了但接线错了、缺诊断入口，还是真的缺功能。这个结果只是起点；动手前仍然要查仓库、测试、工具和文档。
The installer adds the workflow skill where skills are supported, registers time-library MCP where the platform supports MCP, and keeps backup/receipt records for local config writes.
Quick Install
2026.8.13 is the current published release. Use the versioned installers below.
macOS / Linux:
curl -fL -o time-library-install.sh https://github.com/strmforge/time-library/releases/download/v2026.8.13/install.sh
bash time-library-install.sh
Windows PowerShell:
iwr https://github.com/strmforge/time-library/releases/download/v2026.8.13/install.ps1 -OutFile .\install.ps1
.\install.ps1
If you downloaded the release zip, you can also use the Windows installer entry or the macOS installer entry from the extracted release folder.
Windows installs default to %LOCALAPPDATA%\time-library. To choose a path
before the install:
$env:TIME_LIBRARY_INSTALL_DIR = "D:\Apps\time-library"
iwr https://github.com/strmforge/time-library/releases/download/v2026.8.13/install.ps1 -OutFile .\install.ps1
.\install.ps1
If you already downloaded the repo, you can also run:
.\install.ps1 -Dir "D:\Apps\time-library"
WSL is only for development or advanced testing. Normal Windows installs should use the Windows PowerShell command above.
On Windows, use the Time Library tray icon after install. On macOS, use the Time Library menu bar icon. Both can open the local console, show health, and catch up missed records.
You can also open the local console directly:
http://127.0.0.1:9850
Safe First Check
For install checks, do not use /memory first. It may run real recall. Ask the client to call time_library_recall with:
{"query":"capability check","mode":"capability_check"}
A good first result should include:
read_only: true
recall_performed: false
raw_excerpt_returned: false
mcp_tools: ["time_library_recall"]
Only run real recall after you explicitly choose to test memory retrieval.
Record Doctor
To check whether records are guarded before testing recall, run:
python3 tools/record_doctor.py
It prints a short read-only report for source records, raw mirrors, the canonical index, and memory/experience links. It does not run recall, backfill, model calls, or platform writes.
Local Diagnostics
Diagnostics are useful, but they should not become the daily path. Time Library keeps health checks, record checks, and troubleshooting reports separate from ordinary recall so a diagnostic job does not overload the workstation that is also running your local agents.
Use Record Doctor and the local health page first. Deeper evaluation work should run in a separate maintainer workspace and should not be treated as a product feature or public leaderboard claim.
What The Local Page Shows
Open http://127.0.0.1:9850 to see:
- which AI tools are present on this machine;
- which ones can run a safe capability check;
- which ones are already connected or ready for local AI tool integration;
- whether source records, raw mirrors, the canonical index, and memory/experience links are guarded;
- whether a tool looks recently used or has been quiet for a while;
- where new raw records are being stored.
On Windows and macOS, the tray/menu bar icon gives you the same entry point without remembering the port. The local watcher keeps running and can backfill missed records after restart or repair.
Supported local AI tool entries can be connected automatically. Conversation import uses verified local formats, and capability check remains no-recall until an agent calls real recall.
What Makes It Different
- Source-backed memory: recall can carry
source_refs, raw excerpts, library ids, and rank reasons. - Preference and work-experience shelves: The preference shelf keeps preference and intent experience; The work-experience shelf keeps work experience and validation paths. Experience is not a skill library.
- Read-only pre-work checks: agents can check existing context before they edit, so a finished feature does not get rebuilt just because the next window forgot it.
- Explicit memory authority: passive capture, recall, context injection, direct answers, and platform actions are separate levels. OpenClaw-style interception is passive by default.
- Evidence-bound model use: model calls are optional and must answer from supplied evidence with supporting refs or return
UNKNOWN. - Traceable experience evolution: candidates, review queues, validation receipts, apply gates, adoption receipts, and rollback overlays keep useful work paths improving while preserving raw records and receipts.
- Record doctor: a one-click self-check shows whether source records, raw mirrors, the canonical index, and memory/experience links are guarded.
- A timeline you can trace back: different tools leave different clues, but Time Library keeps them in one source-backed timeline. Raw records stay first; useful experience can settle into preference records, work-experience records, toolbooks, or errata with source refs, collection ids, lifecycle state, and receipts.
- Organized local records: new records are grouped by computer first, then by the AI tool that produced them, so a multi-device setup can stay understandable.
- Claude is handled carefully: Claude Desktop and Claude Code CLI can both connect, but they remain separate surfaces. Official, relay, and CLI-related records keep attribution boundaries.
- Hermes can inspect sources itself: Time Library can provide raw/source-ref pointers and observe native feedback, while Hermes-owned skill changes remain Hermes-owned.
Current Published Release: 2026.8.13
2026.8.13 strengthens source-preserving record reliability, bounded recoverability
diagnostics, and unattended Guardian task handling while retaining the
fail-closed, relay-preserving Codex MCP guard and watcher from 2026.7.25. The
local setup, front-door, archive, and transparency boundaries from the
published 2026.7.18 release remain in place. Legacy memcore-cloud roots
remain migration and uninstall fallbacks so existing local data is preserved.
See RELEASE_NOTES_2026.8.13.md for this release, UPDATE_HISTORY.md for older highlights, and CHANGELOG.md for lower-level changes.
AI Tool Surfaces
- Claude Desktop: can use Time Library through local MCP / Desktop Extensions; source records use verified local format collectors.
- Claude Code CLI: can use MCP while staying separate from Claude Desktop.
- Codex: can use the shared skill and MCP entry, and local sessions can become source-backed records.
- OpenClaw: can receive memory support through local entry points, but normal chat is not taken over by Time Library by default.
- Hermes: can consume raw/source-ref pointers and produce native feedback without Time Library writing Hermes skills.
- Other local AI tools: can be recognized from local settings, app folders, package managers, and workspace markers; supported local entries can be connected automatically, and tools are promoted to memory sources once their local formats are verified.
Honest Status (What We Do and Don't Claim)
We would rather under-promise. Proven today: source-exact recall, hybrid search, the multi-agent reading room, borrow records, catalog-on-startup, and a model switch that changes default recall routing.
We do not claim:
- that cross-machine sync is finished — it is at design-audit / partial-remote-source;
- that Hermes auto-adopts production experience, or that an unattended background tick has been observed firing the entire chain — the controlled run-once is proven and the bounded scheduler is implemented, while production adoption remains a separate, closed gate;
- that one vector model is universally better — the switch changes default recall routing, and the current Granite setup has only a local shadow evaluation, not a public benchmark;
- fully autonomous self-training — experience evolution is curated, with review and receipts.
Every change ships in separate proof layers — source/test, local preview, installed runtime, cross-machine, release — signed separately, so "it clicks" is never mistaken for "it works."
Documentation
Uninstall
macOS / Linux:
~/Library/Application Support/time-library/uninstall.sh
Windows:
.\uninstall.ps1
Uninstalling removes the app files only. Local data such as memory/, raw/, config/, and logs/ is kept.
