我们可能从一开始，就把 Agent Memory 想得太复杂了

最近一篇论文《Is Complex Agentic Memory Always Needed?》做了一个很有意思的实验：

不做 Memory Extraction，不做复杂的知识图谱，也不让 LLM 在写入时反复总结、更新记忆

只保留原始对话历史，在需要的时候检索出来，交给 LLM 自己推理

结果呢？

在 LoCoMo、LongMemEval 等四个 benchmark 的 7 组测试设置中，这套简单的 RAW-RAG 方案拿下了 6 组第一，最多领先对比的最强 Agentic Memory 系统 5.53 分

这恰好也是我们开发 Obelisk 时一直坚持的思路

我们认为，Agent 不应该只能看到别人替它总结好的「记忆」

因为一次代码修改背后的原因、一次失败的尝试、几个决策之间的依赖关系，都可能在总结和信息提取的过程中丢失

与其猜测未来什么信息值得记住，不如保留完整历史，让 Agent 在真正需要的时候自己去找

所以我们做了 Obelisk：

一个开源的 Coding Agent Session 检索系统

它可以直接查询 Claude Code、Codex、Hermes、Pi、Opencode 等 Agent 的历史 Session，包括对话、工具调用、Subagent 和工作流。

不依赖 Embedding，也不需要预先从历史中提取 Memory 基于 SQLite FTS5 和 Agent 主动查询，让 Agent 能够重新找到过去的原始证据

比如：

• 为什么当初要这样设计这段代码？
• 上一个 Agent 的任务进行到哪里了？
• 我们之前尝试过哪些方案，为什么失败？
• 这个项目的架构是如何一步步演变的？

Memory 不应该是一份不断被改写的摘要，而应该是一段随时可以追溯的历史

也许下一代 Agent Memory 的关键，不是让 Agent 更努力地记住，而是让它拥有找回过去的能力

Obelisk 完全开源： https://github.com/tommy0103/obelisk…

论文：https://arxiv.org/abs/2610.30755

tommy0103/obelisk

Source: https://github.com/tommy0103/obelisk

Past Claude Code, Codex, GitHub Copilot, DeepSeek Harness, Hermes Agent, Kiro, Kimi Code, OMP, OpenCode, Pi, and ZCode sessions – queryable by your agent, browsable by you.

English · 中文

Obelisk indexes your local coding-agent history into one SQLite database and serves it to two audiences:

• Agents — the obelisk CLI plus an agent skill that teaches coding agents to search their own session history. The agent writes JS queries, runs them locally, and answers in plain language.
• Humans — an Electron desktop app to browse sessions, manage memories, view usage stats, and see weekly recap cards.

Both read the same ~/.obelisk/obelisk.sqlite, kept fresh by watching every configured provider directory.

20K lines of scattered JSONL → something your agent can search() and sql() against in milliseconds.

Use it from your agent

/obelisk 上次 auth bug 最后到底改了哪些文件，为什么这么改
/obelisk 这个文件最近在哪些 sessions 里被反复修改
/obelisk 找出最近失败的 tool calls，它们分别发生在哪些任务里
/obelisk 那个 review workflow 的 subagents 各自结论是什么
/obelisk recap this week

Install

Install with your agent (recommended)

The shortest path is to give the bootstrap guide directly to a coding agent with shell access. Paste this as a prompt into Claude Code, Codex, or another agent — not into your terminal:

Install Obelisk by fetching and following this guide:
curl -fsSL https://raw.githubusercontent.com/tommy0103/obelisk/main/SKILL.md

The agent will ask before changing your machine, install and verify the CLI, then ask whether the formal /obelisk skill should be installed for the current project or globally. The bootstrap guide is only for one-time setup; it is not the query skill itself.

Install manually

Obelisk requires Node.js 22.13 or newer. Install the platform-neutral CLI (zero runtime npm dependencies — it uses Node’s built-in SQLite):

npm install --global @obelisk-apps/cli
obelisk --version

On macOS, Linux, or WSL, the CLI-only installer is equivalent:

curl -fsSL https://raw.githubusercontent.com/tommy0103/obelisk/main/install.sh | sh

Then install the agent skill:

obelisk install

obelisk install delegates to the standard skills installer for tommy0103/obelisk-skill. Then in any agent session: /obelisk <your question>. First run builds the index (~5 seconds for 100 sessions); after that it rebuilds incrementally.

How it works

You ask a question
  ↓
Agent writes a JS query against the SQLite index
  ↓
Runs it via obelisk --query <script>
  ↓
Reads the JSON result, answers in natural language

Default query surface (CLI 0.3.0+): overview(), sessions(), search(), messages(), memories(), summaries(), and sql(). Specialized helpers and legacy script globals remain available. Signatures, options, and query patterns live in the API reference.

Memory layer

When a retrieval produces a conclusion worth keeping, the agent proposes a markdown memory file. After your approval, it registers the file with obelisk --attune <script>. Memories are recalled via memories() in future sessions — a synthesis cache, not a replacement for raw evidence.

Recap

The optional /obelisk recap flow turns a period of sessions into shareable weekly/monthly recap cards. It loads its references only for explicit recap intent, starting at skill-doc/references/recap/overview.md and proceeding card by card.

Browse it in the app

• Sessions — browse all sessions with search, project filtering, readable tool calls (diffs, terminal output, file viewers)
• Memory — list and detail views for registered memory files
• Activity — GitHub-style heatmap, weekly/cumulative token charts
• Recap — shareable weekly/monthly recap cards with archetype theming
• Settings — data source configuration, auto-refresh, rebuild index

Prebuilt releases are available for macOS and Linux amd64 from Releases, with in-app updates (Sparkle on macOS, DebUpdater on Debian installations). On macOS, Windows, and Linux you can also run the app from source — Node.js 22.13+ for the build tools; the app itself runs on Electron’s embedded Node:

git clone https://github.com/tommy0103/obelisk.git
cd obelisk/app
npm ci
npm run dev

Debugging, packaging, and the release workflow are maintainer docs; they live in CONTRIBUTING.md.

Note: while the desktop app is running it owns index writes, so CLI invocations stay read-only.

What gets indexed

Every provider projects into the same shared layers; a few layers exist only where the source tool has the concept:

Full-text search via FTS5 covers all layers. Per-provider source locations are listed under Provider coverage.

Provider coverage

Every provider indexes into the same schema; rows carry a source value and non-Claude IDs are provider-prefixed so they cannot collide.

Where a provider attests superseded history, Obelisk preserves it as inactive rows that normal queries omit; supported query helpers can opt in with includeInactive: true.

Non-default locations — a Pi --session-dir, an OMP custom root, an OpenCode data root (an explicit configured root overrides $OPENCODE_DB), the VS Code Insiders User directory — can be selected in the app’s Settings.

Adapter internals (identity hashing, supersession replay, per-provider root discovery) are documented in retrieval semantics and the ADRs.

Contributing

Contributions are welcome. Read CONTRIBUTING.md before opening a PR — it is short, and it is written from what actually blocked past PRs rather than from generic style rules. The parts worth knowing up front:

• Run every claim in your PR description end to end. The most common reason a PR stalls here is a capability that is advertised but unreachable — including inputs shown in screenshots.
• Assert the requirement, not the implementation. Copy the sentence from the issue into your test name.
• Transcript content is attacker-controlled. Obelisk indexes third-party agent logs; anything reaching shell.*, fs.*, innerHTML, or DDL is deny-by-default.
• Re-run verification after merging main. A merge voids every result above it, including your own noted limitations.

Star History

License

Copyright (C) 2026 tommy0103 and contributors.

Obelisk is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0-only); see LICENSE. Derivative works are welcome: if you distribute a modified version, please keep the per-file copyright notices intact and mark your modifications prominently with a date, as AGPL-3.0 §5 requires.