pi-obsidian-memory

Long-term memory for the pi coding agent, stored in Obsidian vaults

Packages

Package details

extensionskill

Install pi-obsidian-memory from npm and Pi will load the resources declared by the package manifest.

$ pi install npm:pi-obsidian-memory
Package
pi-obsidian-memory
Version
0.2.0
Published
Oct 26, 2026
Downloads
668/mo · 254/wk
Author
davy121
License
MIT
Types
extension, skill
Size
770.8 KB
Dependencies
5 dependencies · 2 peers
Pi manifest JSON
{
  "skills": [
    "./skills"
  ],
  "extensions": [
    "./src/pi/extension.ts"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

pi-obsidian-memory

Long-term memory for the pi coding agent, stored as Markdown notes in Obsidian vaults.

pi forgets everything when a session ends. With this package it keeps what matters (your preferences, the rules and decisions of each project, lessons from past mistakes) and brings it back when it helps. You can read and edit every note in Obsidian.

Contents

  1. Requirements
  2. Install
  3. Set up
  4. Coming from pi-observational-memory
  5. How it works
  6. Settings
  7. Where things are stored
  8. Privacy
  9. Troubleshooting
  10. Known limits
  11. Development

Requirements

What Why Required?
pi Runs the package. Tested with pi 1.3.0. Yes
Node 22.19 or later pi runs the package on it. Yes
A chat model configured in pi Turns the conversation into observations for compaction (observer) and finds notes for each message (memory agent). Any model pi can use works; a small, cheap one is enough for the memory agent. Yes
A TypeSafe API key (TYPESAFE_API_KEY) Jev, TypeSafe's classifier, decides which notes help each message and checks shell commands (see Guards). Get a key at typesafe.ai. Price: $0.042 per million input tokens. Strongly recommended
Obsidian Reading and editing the notes. No: notes are plain Markdown

Without TYPESAFE_API_KEY: no notes are added to your messages (pi tells you once per session), and every bash or powershell call is refused, because the shell-edit guard cannot check it.

Setting the TypeSafe key

Set the environment variable before starting pi:

# macOS / Linux (add it to ~/.bashrc or ~/.zshrc to keep it)
export TYPESAFE_API_KEY=your-key
# Windows (kept for new terminals)
setx TYPESAFE_API_KEY "your-key"

The key can also be pi's stored credential for the typesafe provider. It is never written to this package's config.

Install

pi install npm:pi-obsidian-memory

Then restart pi. To try it for one run without installing: pi -e npm:pi-obsidian-memory.

pi list shows installed packages; pi update updates them.

Set up

  1. Create the global vault. Run /memory-setup in pi. It asks:

    • where the global vault goes: a new folder, or an existing folder (an Obsidian vault you already have works);
    • which chat model runs the observer, which turns the conversation into observations for compaction; the model in use is the default;
    • which chat model runs the memory agent; the list is sorted by price, cheapest first.

    It shows what each choice sends where, then writes the config and the vault.

  2. Give each project its own vault. Open pi in the project folder and run /memory-init. A folder without a project vault uses the global vault only.

  3. Use pi as usual. Say "remember that …" to save something. Until the next release, nothing else is saved to the vaults.

Run /memory-setup again at any time to open the settings screen. Both commands need pi's interactive mode.

Coming from pi-observational-memory

This package now observes and compacts the session the way pi-observational-memory does. Both packages compact, and pi keeps only one compaction, so uninstall pi-observational-memory:

  1. Run pi list to see its source.
  2. Run pi remove <source> with that source.
  3. Restart pi.

/memory-doctor warns while it is still loaded. Its files in .memory/ are left as they are. Importing them into the vaults arrives with the next release.

How it works

Two vaults

  • Global vault: about you. How to address you, your preferences, your patterns, lessons that apply everywhere.
  • Project vault: about one project. Its rules, decisions, facts, open issues.

Each note holds one point. Its file name is the statement (for example pnpm/Use pnpm, never npm.md), the body has the detail, and the frontmatter says what kind of memory it is:

---
id: 01JABCDE7XKQ2M9V3RTN8WPZ5F
type: rule
created: 2026-10-04
updated: 2026-10-04
dontForget: true
---

Use pnpm for every install and script. In CI: `pnpm install --frozen-lockfile`.

Types: rule, procedure, constraint, fact, issue, question, decision, lesson, goal, risk, idea, reference. Notes are written in English. They are never deleted: a retired note moves to archive/. The full format is in docs/vault-format.md.

What happens during a session

  1. Session start. Every note marked dontForget: true is loaded into pi's system prompt, in full.
  2. Each message. Jev decides whether stored memory would help. If yes, the memory agent searches and reads the vaults, Jev scores each note it found, and the notes that help are added after your message, within context.budgetChars (12,000 characters by default). A note you changed since it was loaded is sent again.
  3. Observer and compaction. After each turn, the observer reads each new stretch of the conversation (10,000 tokens by default) and writes dated observations into the pi session. When the context reaches 150,000 tokens, memory compacts the session: the older part is replaced by a block with the memory map (the names of the notes in both vaults) and the observations, and the last 20,000 tokens are kept word for word. /compact and pi's own compactions use the same block. A run that was compacted while it used tools goes on by itself. Observations stay in the session: until the next release, only what you ask pi to remember is saved to the vaults. The cycle follows pi-observational-memory.

You change notes in Obsidian whenever you want. pi must read a note before changing it, and a note you edited since then is never overwritten.

Tools the agent uses

Tool Does
memory_search Finds notes by words and by meaning, in the global vault, the project vault or both
memory_get Reads a note, a folder's index, or anything under archive/
memory_add Saves a new note
memory_update Changes a note
memory_archive Moves a note to archive/
memory_succeed Replaces a note with a new one; both link to each other

The bundled skill obsidian-memory teaches the agent where notes go and how to use these tools.

Commands

Command Does
/memory-setup First setup; afterwards, the settings screen
/memory-init Gives the current project its own vault
/memory-doctor Checks the vaults, the config, the models and the keys, and says what is wrong
/memory-export Saves both vaults to a new .zip

Guards

  • Vault guard. The agent reaches the vaults only through the memory tools. pi's file tools (read, write, edit, grep, find, ls) on a vault are refused, and so are shell commands that write a vault path out. You and Obsidian are not affected.
  • Shell-edit guard. Before every bash or powershell call, Jev checks whether the command edits a file: a redirect, an in-place edit, inline code, a script that writes, a copy or download that creates a file, or a hardlink. If so, the call is refused and the agent is told to use edit or write. When Jev cannot answer, the call is refused.

Dashboard

Each vault gets a Memory.base file: an Obsidian view of all notes, the Don't forget notes and the archive. Your edits to it are kept.

Settings

All settings are changed in the /memory-setup screen, or by asking pi. You never need to edit the file.

Setting Default What it does
Global vault chosen at setup Folder of the global vault
Projects none Each project folder and its vault, added by /memory-init
Observer model the model in use at setup Chat model that turns the conversation into observations
observer.thinking low Thinking level of the observer model
observer.chunkTokens 10,000 Tokens of conversation the observer reads at a time
observer.concurrency 4 Observers that run at the same time
compaction.atContextTokens 150,000 Context tokens at which memory compacts the session
compaction.tailTokens 20,000 Tokens of the latest conversation kept word for word after a compaction
compaction.mapTokens 2,000 Tokens of the memory map in the compaction block
Memory agent model chosen at setup Chat model that searches and reads notes for each message
context.budgetChars 12,000 Characters of relevant notes added at session start and to each message (Don't forget notes do not count)
limits.maxStatementChars 100 Longest note name
limits.maxBodyChars 1,000 Longest note body
limits.maxNotesPerFolder 20 /memory-doctor suggests splitting a folder above this
Embeddings local Model for search by meaning: local (no key), OpenAI (OPENAI_API_KEY) or Google (GEMINI_API_KEY)
Decision model Jev Jev (TypeSafe, cutoffs 0.5/0.5) or Laya (local server, you set the cutoffs)

Every field, the file format and running Laya locally are in docs/config.md.

Where things are stored

What Where
Notes The vault folders you chose
Settings pi-obsidian-memory/config.json in pi's agent folder (~/.pi/agent by default)
Search index pi-obsidian-memory/index/ in pi's agent folder; rebuilt from the notes when needed
Local search model pi-obsidian-memory/models/ in pi's agent folder; downloaded once, about 135 MB

The vaults hold only notes, their index.md files, _templates/ and Memory.base, so they can live in any synced folder (OneDrive, iCloud, Dropbox).

Privacy

  • The observer and the memory agent send conversation and note content to the chat models you choose.
  • Observations are kept in the pi session file, like the conversation.
  • Jev sends your messages, candidate notes and shell commands to TypeSafe.
  • Search by meaning runs on your machine by default; OpenAI or Google receive note content and queries only if you choose them.
  • /memory-setup shows each of these before it saves.
  • Notes are saved as written, without masking. The config file never holds a key.

Troubleshooting

Run /memory-doctor first: it checks the vaults, the config, the models and the keys, and says what is wrong. The only thing it offers to change is a blank note, after asking you.

You see Do
"Memory is off" or "Run /memory-setup first" Run /memory-setup
"No chat model is available" during setup Connect a model in pi (/login), then run /memory-setup again
No relevant notes are ever added Check that TYPESAFE_API_KEY is set, then run /memory-doctor
"Refused: the shell-edit guard could not check this command" Same: the shell-edit guard needs Jev
"Memory observer failed: …" The text after the colon names the cause, for example a model without a key. It tries again a minute later; /memory-doctor checks the observer model
"… so pi compacted it on its own" The observer had not covered the start of the session yet. Nothing to do: it catches up, and the next compaction uses its observations
"Memory could not compact the session: …" The text after the colon names the cause, for example "Nothing to compact (session too small)". Nothing was compacted, and a run the compaction interrupted goes on. Nothing to do: memory tries again once the conversation has grown
No API key found for "<provider>" in /memory-doctor Log in to that provider in pi (/login) or set its key

Known limits

  • Paths are compared ignoring case on Windows and macOS and with case on Linux, whatever the volume.
  • The vault guard does not see a vault path the shell builds indirectly (variables, globs, scripts), tools of other extensions, or foreground subagents.
  • Until the next release, observations are not turned into notes: the vaults get only what you ask pi to remember.
  • With pi-observational-memory also installed, only one of the two compactions is kept (see Coming from pi-observational-memory).

Development

npm install
npm run lint
npm run typecheck
npm test              # offline: pure tests and tests on a real temporary disk
npm run test:e2e      # the real pi over RPC, through OmniRoute
npm run test:model    # live tests: real models and network
pi -e ./src/pi/extension.ts

npm run test:e2e runs one test per feature with the real cc/claude-haiku-5-5 through OmniRoute (pi-omniroute-sync) and the real Jev. It reads OMNIROUTE_URL, OMNIROUTE_API_KEY and TYPESAFE_API_KEY from the environment. CI runs it on Windows, and lint, typecheck and npm test on Windows, macOS and Linux.

Releases are automatic: a push to main with a feat or fix commit bumps the version, rewrites CHANGELOG.md, publishes to npm and creates the GitHub release.

More: vision, roadmap, note format, settings.

License

MIT