Repository avatar
Other Tools
v0.1.3
active

orimem

io.github.Wen5555/orimem

Long-term memory for coding agents in your own Obsidian vault: recall, write-back, capture.

Documentation

OriMem icon

OriMem

Long-term memory for coding agents, kept in your own Obsidian vault.

One local MCP server and one skill, shared by Claude Code, Codex and DeepSeek Harness — recall project context across sessions, write confirmed results back with receipts, and capture every conversation automatically.

CI Python MCP Platforms License: MIT

English · 简体中文


Why OriMem

Coding agents forget everything when a session ends, and each agent keeps its history in its own private format. OriMem gives them one shared, human-readable memory:

  • 🗂️ Your vault is the memory. Plain Markdown in an Obsidian vault you can read, edit and back up. The index is just a cache.
  • 🔎 Hybrid recall. Multilingual E5 embeddings + SQLite FTS5 + wiki-links. Short snippets first, exact evidence on demand, so context stays small.
  • ✅ Verified write-back. prepare_record → diff → commit_record with hash checks, backups, read-back and receipts. It never silently overwrites a note you edited.
  • 🎙️ Automatic capture. Hooks save each Claude Code / Codex / dsh session as a transcript note when a turn ends. Tool output, reasoning, injected system context, sub-agents and automations are left out.
  • 📅 Daily summary (opt-in). The agent asks you once; if you agree, a scheduled job writes the day's journal through the same verified writer.
  • 🔒 Local only. No cloud calls, no telemetry. The embedding model is pinned and checksum-verified, and runs on CPU through ONNX Runtime (no PyTorch needed).

Quick start

Let your agent install it (recommended). Paste this into Claude Code, Codex, DeepSeek Harness or any coding agent:

Install and set up OriMem for me by following https://raw.githubusercontent.com/Wen5555/OriMem/main/llms-install.md

The agent asks you four questions (vault folder, folder language, capture, search mode), installs without further prompts, connects every agent it finds, verifies with orimem doctor and tells you what to restart. If OriMem is already set up (for example from another agent), it skips all of that and only connects itself. The steps it follows are in llms-install.md.

Or run the installer yourself:

Windows (PowerShell)

irm https://raw.githubusercontent.com/Wen5555/OriMem/main/scripts/install.ps1 | iex

macOS / Linux

curl -LsSf https://raw.githubusercontent.com/Wen5555/OriMem/main/scripts/install.sh | sh

The installer sets up uv if needed, installs OriMem into an isolated environment (about 130 MB, Python provided by uv), then runs orimem init. init asks a few questions — vault path, folder language (Chinese or English, from your system locale), conversation capture (default on), daily summary (default "ask me later") — downloads the 490 MB embedding model, and connects every agent it finds. Restart your agents afterwards.

No Obsidian yet? Not required: the vault is a plain Markdown folder. init prepares it for a good first open (welcome note, new notes to the inbox, daily notes to the daily folder) and tells you how to install or add it; orimem open opens it later. Short on disk or offline? orimem init --lexical skips the model and uses full-text search only (switch back any time with orimem config set recall.mode hybrid).

WhatSize
OriMem environment (onnxruntime, numpy, tokenizers, …)~130 MB
Embedding model (model.onnx + tokenizer)~490 MB (none with --lexical)
uv + Python 3.12, only if not already installed~70 MB + ~70 MB
Typical total~0.6 GB (~0.75 GB on a bare machine; ~0.13 GB lexical-only)
Optional PyTorch backend (orimem-mcp[torch])+1 GB or more

Then just talk to your agent:

"What did we decide about the export feature last week?" "Record that the WebGPU renderer is done and the next step is drag-to-reorder layers."

Using several agents

OriMem is installed once per computer. Claude Code, Codex and DeepSeek Harness all share the same vault, embedding model, index and settings, so adding an agent never re-downloads or re-asks anything.

SituationWhat to do
The agents are already installed when you set OriMem upNothing: init connects every agent it finds.
You install another agent later (e.g. OriMem set up from Codex, dsh installed next week)Nothing: the next time any agent uses OriMem, the new agent is connected in the background (agents.auto_connect). Restart it once.
You want it connected right noworimem install --agent dsh (or claude, codex), or paste the one-line prompt above into that agent: it sees OriMem is already set up and only connects itself.
You removed an agent on purposeorimem uninstall --agent codex; it is remembered and never reconnected automatically. orimem install --agent codex brings it back.
Check what is connectedorimem doctor lists connected agents, agents found but not connected, and any step an agent still needs.

One-time steps per agent:

  • Codex runs new hooks only after you approve them: on the next codex start accept "Hooks need review" (or open /hooks and press t). Until then the MCP tools work and conversations are still captured every 2 minutes, but project context is not injected. Approve again if Codex says the hooks changed after an upgrade.
  • DeepSeek Harness installed from npm (@deepseek-ai/dsh) does not ship the hook bridge; orimem install adds it to each profile through dsh plugin, which needs pnpm (npm install -g pnpm). The desktop app already includes it.
  • Claude Code can use the plugin instead of orimem install (below); OriMem then leaves Claude's own settings alone.
Other ways to install
uv tool install "orimem-mcp>=1.0.0rc2"
pipx install "orimem-mcp>=1.0.0rc2"

The >=1.0.0rc2 part makes uv/pip take the newest OriMem, release candidates included (other packages stay on stable releases); the installers above do the same.

The PyPI distribution is orimem-mcp; the command it installs is orimem. To run the MCP server without installing it first, use uvx --from orimem-mcp orimem mcp.

orimem init

Non-interactive: orimem init --yes --vault ~/Notes --agents claude,codex. Behind a slow network, the installer switches to mirrors automatically when PyPI is unreachable (ORIMEM_MIRROR=cn forces it).

Supported agents

AgentMCP serverSkillHooks (capture on turn end)Transcripts read from
Claude Codeclaude mcp add -s user orimem~/.claude/skills/orimem~/.claude/settings.json~/.claude/projects/*/<uuid>.jsonl
Codex[mcp_servers.orimem] in ~/.codex/config.toml~/.codex/skills/orimem~/.codex/hooks.json~/.codex/sessions/**/rollout-*.jsonl
DeepSeek Harness (dsh)dsh-mcp-client entry in the profile's cordis.patch.yml~/.dsh/skills/orimemdsh-hooks-claude-code bridge~/.dsh/sessions/**/session.v4.jsonl.zstd
Any other MCP clientorimem snippet prints the JSONcopy the skill folder——

Claude Code users can also install it as a plugin (MCP server, skill and hooks in one step; still needs the orimem command):

/plugin install orimem --marketplace Wen5555/OriMem

On older Claude Code versions: /plugin marketplace add Wen5555/OriMem, then /plugin install orimem@orimem.

DeepSeek Harness users can also add the dsh-orimem bundle from the community plugin market (or dsh plugin --profile <name> add dsh-orimem-<version>.tgz from a release); it needs the orimem command from the quick start.

orimem install and orimem uninstall only add or remove OriMem's own entries, and every edited file is backed up first. Marker blocks in TOML/YAML files are removed byte for byte; JSON files keep all other content.

How it works

flowchart LR
    subgraph Agents
        CC[Claude Code]
        CX[Codex]
        DS[DeepSeek Harness]
    end
    CC & CX & DS -- "MCP (stdio)" --> M[orimem mcp]
    CC & CX & DS -. "Stop hook" .-> H[orimem hook]
    M --> S[(local service<br/>loopback only)]
    H -- "capture this session" --> S
    S --> I[(SQLite: chunks · E5 vectors<br/>FTS5 · links · ledger)]
    S <--> V[/"Obsidian vault<br/>(Markdown = source of truth)"/]
    S -- "every 2 min" --> T[agent transcripts]
  • Recall: task_context binds the current project; recall returns ranked snippets; evidence fetches exact source text with offsets and hashes.
  • Write-back: the agent proposes a complete note change; OriMem checks the expected hash, backs up, writes, re-reads, verifies the claimed text and issues a receipt.
  • Capture: per-agent parsers mirror ordinary user/assistant turns into one pipeline, which stages them and projects stable transcript notes into the vault's evidence folder. The Stop hook only hands the session to the running service (falling back to a detached capture process) and is registered as async for Claude Code and Codex, so it never delays a turn; dsh runs hooks synchronously, where it returns in a fraction of a second. Context hooks (session start, prompt submit) stay synchronous because their output is the injected project context. Every parse records the client/format version and checks the transcript shape; unsupported formats are not captured and drift shows up in orimem doctor.

Configuration

Settings live in ~/.orimem/config.toml. Change them with orimem config set <key> <value> or ask your agent, which uses the config_set tool.

KeyDefaultMeaning
vault~/.orimem/vaultObsidian vault path
capture.enabledtrueCapture conversations into the vault
capture.sources["claude","codex","dsh"]Which agents to capture
capture.sinceinstall timeOlder sessions are not back-filled (orimem capture --backfill does it on request)
capture.exclude_cwd[]Never capture sessions started in these folders
daily_summary.enabled"ask""ask" = ask once, off until answered · true · false
daily_summary.time"22:30"When the daily job runs
daily_summary.agentautoclaude or codex runs the summary
recall.mode"hybrid"hybrid = vectors + full text (needs the model) · lexical = full text and links only, no model
recall.hf_endpointautoModel download endpoint (falls back to hf-mirror.com)
mcp.tools"core"core publishes 8 everyday tools · all adds the advanced ones
mcp.extra[]With core: individual advanced tools to publish, e.g. ["health"]
agents.auto_connecttrueConnect agents installed after OriMem (they share the vault, model and settings)
projects.link"auto"auto: a new project folder (git checkout or build manifest) gets its own project page automatically · offer: the agent suggests it once · off
projects.import_historytrueAuto-linking also imports that folder's conversations from before the install
projects.recordtrueIn a linked project, agents keep the project page up to date when a session delivers something
profile.sharetrueShare the vault-root profile note with every agent at session start (profile.max_chars, default 1500)
layout.languagesystem localezh or en folder preset
layout.<role>presetOverride one folder: inbox, drafts, projects, courses, areas, knowledge, daily, weekly, archive, evidence, legacy, workspace; also course_home, weekly_note, knowledge_index
layout.extra_writable[]Further folders automatic write-back may touch
maintenance.backup_days30Keep write-back backups and finished batch records this long
maintenance.projection_backup_days3Backups of auto-generated transcript notes are kept only briefly

Reference

CLI
CommandWhat it does
orimem install / uninstall [--agent claude,codex,dsh] [--dsh-profile NAME]Connect or disconnect agents
orimem init [-y] [--lang zh|en] [--lexical] [--json]First-time setup (re-running it is safe)
orimem doctor [--json]One-screen health check: vault, index, search, agents, capture, Obsidian, storage
orimem prune [--dry-run]Delete old write-back backups and finished batch records (also runs daily)
orimem openOpen the vault in Obsidian (or say how to install/add it)
orimem onboard scan / import --cwd DIR [--days N] / statusBring conversations from before the install in, by working folder
orimem link [--cwd DIR]Create or bind the project page for a folder (the hook does this automatically)
orimem capture [--source X] [--session ID] [--backfill]Capture now
orimem sync · orimem recall "query" --scope "10 项目"Index / search from the shell
orimem daily status / run [--dry-run] / applyDaily summary
orimem model status / download [--torch] / import DIREmbedding model
orimem config get / set / pathSettings
orimem snippetMCP JSON + AGENTS.md snippet for other clients
orimem --home DIR …Use another data directory (written into generated agent commands)
MCP tools

Published by default (8):

ToolPurpose
task_contextBind the current folder/task to a project entry
recallHybrid (or lexical) search, snippets
evidenceExact source text for result ids
initializeCreate or bind a project/course entry to a folder (checks for duplicates; several folders per project)
list_sessions · get_session_messagesBrowse captured conversations
prepare_record · commit_recordVerified write-back

config_set is also published while the daily summary is still "ask", so the agent can record the answer.

Advanced (orimem config set mcp.tools all, or list some in mcp.extra):

ToolPurpose
lifecycleSource reading, candidates, corrections
health · sync · archive_historyMaintenance
config_get · config_set · capture_nowSettings, capture now
Vault layout

init picks the Chinese or English preset from your locale (--lang to choose); every folder can be renamed with orimem config set layout.<role> "<path>" (existing notes are not moved). orimem config get layout shows the resolved layout.

English preset                    Chinese preset
00 Inbox/ (AI Drafts/)            00 收件箱/ (AI 草稿/)      inbox, knowledge candidates
10 Projects/                      10 项目/                   one page per project
15 Courses/<course>/Course Home   15 课程/<课程>/课程主页
30 Knowledge/                     30 知识/                   reusable knowledge
50 Journal/Daily · Weekly         50 日记/每日 · 每周
90 Archive/Conversations/         90 归档/对话佐证/          captured transcripts
95 AI Workspace/                  95 AI 工作区/              agent work notes

Agents may write projects, courses (homepage only), knowledge, daily/weekly journals, AI drafts and the AI workspace (plus layout.extra_writable). The evidence folder is written only by OriMem's capture, and the profile note only when a plan carries profile_update: true (an explicit request from you). .obsidian is only written once, when init creates a vault Obsidian has never opened; files outside the vault are never touched. Vaults set up by 0.1.x keep 90 归档/Codex佐证.

FAQ

What happens when I upgrade? uv tool upgrade orimem-mcp (or re-run the installer), then restart your agents. Coming from 0.1.x: uv tool install --force "orimem-mcp>=1.0.0rc2" once, since that install was pinned to stable releases. The first time the new version runs, it refreshes its own MCP/skill/hook entries in the agents it is already installed into and adds anything new to the vault (it never overwrites a note). orimem doctor shows the result.

Does OriMem keep growing on disk? Write-back keeps backups and receipts, so a retention policy runs daily: transcript-note backups are kept 3 days, other backups 30 days (maintenance.*). orimem doctor shows the storage used and orimem prune cleans up immediately.

Are secrets removed from captured conversations? No. Transcripts are stored as you typed them, because a rewritten transcript is no longer evidence. Don't paste secrets into agent chats, exclude sensitive folders with capture.exclude_cwd, and keep the vault out of public sync if it holds sensitive work. Tool output and secrets that only appear in tool calls are not captured.

Do I need Obsidian? No. Agents read and write the Markdown folder directly; Obsidian is the optional UI. orimem doctor shows whether it is installed and whether the vault is added.

Can I run without the embedding model? Yes: orimem config set recall.mode lexical (or init --lexical). Recall then uses SQLite full-text search and links only; set hybrid and run orimem model download + orimem sync to get vectors back.

What happens when I start working in a new project? Nothing for you to do. The first session in a project folder (a git checkout or a folder with a build manifest) creates its project page, binds the folder (worktrees and extra clones are recognised), imports that folder's conversations from before the install as evidence, and the agent fills in an overview. If a similarly named entry already exists, the agent asks once whether to bind or create. From then on every session there gets the project context, and agents update the page when a session delivers something. Set projects.link to offer or off to change this. For older history across many folders: orimem onboard scan lists them, orimem onboard import --cwd <folder> brings the chosen ones in.

How do agents share what they know about me? init puts a profile note at the vault root (About Me.md / 个人画像.md). Whatever you write there reaches every agent at session start (Claude Code via the MCP server instructions, Codex and dsh also via the session hook); an empty template is never shared. Agents change it only when you explicitly ask them to remember something about you. Don't put secrets in it: it is sent to whichever model provider each agent uses.

I set OriMem up from Codex. Does dsh (or Claude Code) get it too? Yes. OriMem is installed once per computer and every agent shares the same vault, model and settings. Agents installed later are connected automatically the next time any agent uses OriMem (agents.auto_connect; agents you removed with orimem uninstall --agent X stay removed), or right away with orimem install --agent dsh. Codex asks you once to approve OriMem's hooks (/hooks). With the npm build of dsh, OriMem adds the hook plugin dsh needs (this uses pnpm).

Do the hooks slow my agent down? Capture runs in the background (async Stop hook in Claude Code and Codex). The session-start and prompt hooks take about a quarter of a second because they inject project context.

Do I need a GPU or PyTorch? No. Embeddings run on CPU through ONNX Runtime. orimem[torch] plus orimem model download --torch enables the optional PyTorch backend, which produces identical vectors.

Can I use my existing vault? Yes. Point vault at it. OriMem only adds the folders it needs and never rewrites existing notes outside the verified write-back flow.

I'm in mainland China. Both installers detect an unreachable PyPI and switch to the Tsinghua PyPI mirror, the npmmirror Python builds and hf-mirror.com.

How do I turn capture off for a project? orimem config set capture.exclude_cwd '["D:/work/secret"]'. To turn it off everywhere: orimem config set capture.enabled false.

How do I remove everything? orimem uninstall, then uv tool uninstall orimem-mcp. Your vault and ~/.orimem stay until you delete them.

Roadmap

  • Configurable / English vault layout
  • Full-text-only mode without the model
  • More agents (Cursor, Gemini CLI, OpenCode)
  • Optional quantized ONNX model for low-memory machines

Contributing

git clone https://github.com/Wen5555/OriMem && cd OriMem
uv venv && uv pip install -e ".[test]"
uv run pytest

A manual end-to-end checklist for each agent is in docs/TESTING.md. Issues and pull requests are welcome.

License

MIT

PYPI
orimem-mcp
Install Command
pip install orimem-mcp
Runtime: uvx