sessionmemory

Durable memory for coding agents, one folder of searchable pages per project.

An agent starts every session knowing nothing about the last one. Yesterday's session worked around a trap, rejected the obvious approach for a reason, and found the one flag that makes a library behave. All of it ends with the session, and the next one works it out again from nothing. sessionmemory keeps what a project learned and hands it back when the next session starts. Each project keeps its own memory, and no page is ever shared between projects.

The pages follow the memoryfield format by Cal Paterson, described in his article and defined in the memoryfield spec. Each project's learnings/ folder is one field in that format: a flat directory of markdown pages beside one vector index file. You can export that folder, share it, or read it with any tool that speaks the format.

Requirements

• uv. It installs the CLI, and the plugin runs its hooks through it.
• Python 3.13 or 3.14. uv downloads one when none is installed.
• git. The vault is a git repository, and a project is registered by its git remote.
• Claude Code, to run the plugin. The CLI works on its own without it.

The first search downloads the nomic-embed-text-v1.5 embedding model, about 520MB, and caches it under ~/.cache/sessionmemory/models. Nothing else touches the network.

Install

Installation is three steps: the CLI, the plugin, and a vault for them to share.

1. Install the CLI

The CLI is published on PyPI. Install it as a tool, which puts sessionmemory on your PATH:

uv tool install sessionmemory

To upgrade it later, run uv tool upgrade sessionmemory.

2. Install the Claude Code plugin

In Claude Code, add this repository as a marketplace and install the plugin from it:

/plugin marketplace add natelandau/sessionmemory
/plugin install sessionmemory@sessionmemory

The hooks run the sessionmemory from step 1 when its version is at or past the plugin's own. Keep the two in step: after /plugin update sessionmemory@sessionmemory, run uv tool upgrade sessionmemory as well. A plugin newer than the tool falls back to a copy of the CLI it carries, in a Python environment of its own that the first such session builds.

3. Create a vault

The vault is its own directory. Make it a git repository of its own, so your pages and this code do not share a history. Then initialize it:

mkdir -p ~/repos/my-vault
cd ~/repos/my-vault
git init
sessionmemory init ~/repos/my-vault

sessionmemory init writes the three files a vault needs. A marker in _system/vault.toml identifies the directory as a vault. A .gitignore keeps the derived index out of your history. A README explains the layout to whoever opens the vault later. sessionmemory init never overwrites a file, so you can run it again safely.

Then tell the CLI and the hooks where the vault is. Both read the same two places, in the same order: the SESSIONMEMORY_VAULT environment variable, then vault.root in ~/.claude/sessionmemory.toml. Record the root in the file, since a session launched from a GUI or an IDE does not read your shell profile:

[vault]
root = "~/repos/my-vault"

The variable wins when both are set, so exporting it in one shell points that shell at a different vault without touching the file. The plugin documentation lists every other key that file accepts.

Until a directory holds the marker that sessionmemory init writes, every command refuses to touch it. The refusal protects you. If the root points at your home directory by mistake, the first page written scatters a projects/ tree into it.

Nothing commits the vault on a timer. The plugin commits it when a session starts and again when a session ends, so a page reaches git within the session that wrote it. Pushing that history to a remote stays yours to do.

Note: To bring an existing directory of notes under the CLI, run sessionmemory init --force ~/repos/my-vault once. --force means only that the directory already has contents. Nothing existing is overwritten.

Register a project

A project gets memory when its repository is registered. With the plugin installed, the first session you open in a git repository registers it. The session begins with one line that names the slug:

This repository was registered with the vault as project 'invoice-api'.

The slug comes from the git remote, or from the directory name when the repository has no remote. There is nothing else to choose: no tags, no scope, no note type.

Only a git repository is registered for you. A slug is permanent once pages carry it. A session opened in your home directory or a scratch folder must not leave a project named after it in the vault. To register a directory outside git, or to choose the slug yourself, run the command once:

cd ~/repos/invoice-api
sessionmemory project --register --cwd .

✓ registered 'invoice-api'
  └─ root: ~/repos/invoice-api

From then on, a session that starts in a registered repository receives that project's memory. A session that ends or compacts hands its transcript to a background pass, which records what was worth keeping.

Search and write pages

The CLI does two things. It finds pages by meaning, and it creates pages. Reading and editing a page is a job for your editor or your agent's own tools.

sessionmemory search "stripe event delivered twice" --cwd .

~/repos/my-vault/projects/invoice-api/learnings/stripe-retries-a-webhook-for-72-hours-so-the-handler-must-be-idempotent.md
  Stripe retries a webhook for 72 hours, so the handler must be idempotent
  Stripe redelivers an unacknowledged webhook for up to 72 hours, so the handler records the event id and ignores a repeat.

A result is a path, a title, and a summary. A paraphrase finds the page, because search ranks by meaning and not by words in common. A hit has to stand out from the rest of the project's pages, so a query that nothing answers returns no results rather than the nearest pages. Pass --read to print every hit in full.

sessionmemory new learning \
  --title "Stripe retries a webhook for 72 hours, so the handler must be idempotent" \
  --summary "Stripe redelivers an unacknowledged webhook for up to 72 hours, so the handler records the event id and ignores a repeat." \
  --cwd .

✓ created stripe-retries-a-webhook-for-72-hours-so-the-handler-must-be-idempotent.md
  └─ ~/repos/my-vault/projects/invoice-api/learnings/stripe-retries-a-webhook-for-72-hours-so-the-handler-must-be-idempotent.md

The vault path is shortened to ~ here; the command prints absolute paths.

The command writes the frontmatter and prints the path. Write the body into that file, or pass it with --body-file. The title is what every future session sees at its start, and the summary is what a search result shows. Both state the fact and not the topic.

What a session sees

sessionmemory inject prints the block a session starts with. This is the block for a project holding four learnings, one spec, one plan, one reference doc, and two open backlog items:

## Using this vault

Durable memory for this project lives in a vault of markdown pages. Nothing below is
loaded for you: the titles are what the vault holds, and each is one `sessionmemory search`
away. The project's folder has `learnings/` and `logs/`, searched by meaning, beside
`specs/`, `runbooks/`, `reference/`, and `backlog.md`, which are ordinary files you Read
and Edit.
`sessionmemory project --json` prints every path.

  - Before assuming nothing was written down, search: `sessionmemory search "<words>"`
    prints each hit's path, title, and summary, and `--read` prints every hit's whole
    page in one call. Search with a few distinctive words, such as names, identifiers,
    or error text, rather than a sentence. No hits means nothing is recorded, not that
    the query needs loosening.
  - Past sessions, one page each: `sessionmemory search "<words>" --logs`. Search them
    for why something was decided, what happened the last time an area changed, or
    whether a fix was already tried.
  - Open work: read `backlog.md`. An item is one line under a `## <kind>` heading
    (feat, fix, refactor, perf, docs, test, build, ci), sized S, M, or L:
    `- [S] <imperative description> - <YYYY-MM-DD> [#topic]`. Add one with
    `sessionmemory new backlog --kind <kind> --size <S|M|L> --title "..." --topic <topic> --cwd .`,
    which creates the file or heading when missing. Delete a finished line, and one
    that will never be done, directly; never tick or annotate it. Git history is the
    record of what was finished.
  - Specs: `specs/` holds one design record per feature, named `<date>-<topic>.md`
    and kept after the work ships. Before designing or changing a feature, list the
    directory and read any spec whose name matches, so a decision already made is not
    made again. `sessionmemory new spec --title "..." --cwd .` creates one and prints its
    path. Edit it directly after that.
  - Runbooks: `runbooks/` holds step-by-step procedures to follow by hand, such as
    configuring a service, rotating a credential, or recovering from an outage.
    Whenever you write one, create it with `sessionmemory new runbook --title "..." --cwd .`,
    which prints its path, never in another folder or the repository.
  - Reference: `reference/` holds this project's private docs, such as hosting
    settings, where each credential lives, or a data dictionary, kept out of its
    repository. Any it holds are listed below; before working in an area, read the one
    whose name matches. Keep a doc current by editing it in place. Create one with
    `sessionmemory new reference --title "..." --cwd .`, which prints its path. Public
    documentation stays in the repository.
  - Learnings are captured at session end, not by you mid-session. When the user asks
    to keep one now: `sessionmemory new learning --title "..." --summary "..." --cwd .`
    creates the page and prints the path to write prose into. Title and summary state
    the fact, not the topic. Keep a page under 8KB; more detail is another page.

## What this project knows

  - Invoice numbers come from a Postgres sequence, never from max(id) plus one
  - pytest-asyncio needs asyncio_mode = auto or every async test is skipped
  - Stripe retries a webhook for 72 hours, so the handler must be idempotent
  - The nightly reconciliation job must start after the 02:00 bank feed

## Reference docs

  reference/hosting.md

## Open work

  2 open backlog items
  1 spec in specs/

A page body never enters that block, so its cost grows with the number of pages and not with their length. The titles say what exists. sessionmemory search returns what they say.

Add rules to your CLAUDE.md

The guidance block above tells a session how to read and write the vault. It does not control where other skills and plugins write their files. A planning skill writes its plans and specs to the path it chose for itself, often under the repository's docs/. Rules in a CLAUDE.md file send that work to the vault.

Put the rules in your user file, ~/.claude/CLAUDE.md, so that they apply in every registered project. A project CLAUDE.md is committed with the code. Other contributors read it, and they have no access to your vault.

Copy this block into the file:

## Session memory

- Write every plan, spec, and design document into the sessionmemory vault, including
  the ones a skill or plugin generates. Create them with
  `sessionmemory new spec --title "..." --cwd .` and
  `sessionmemory new plan --title "..." --cwd .`. Never write them under the
  repository's `docs/`, which holds public documentation.
- Write every runbook (a step-by-step procedure to follow by hand) with
  `sessionmemory new runbook --title "..." --cwd .`.
- Keep documentation that describes the code and is safe to publish in the repository.
  Put private documentation about a project, such as hosting settings or where each
  credential lives, in the vault with `sessionmemory new reference --title "..." --cwd .`,
  and edit that file in place afterward.

If you use the CLI without the plugin, no hook injects the guidance block. Add this rule to the same block, so that each session reads the guidance itself:

- At the start of each session in a registered project, run
  `sessionmemory inject --cwd .` and follow the guidance it prints.

Documentation

Development

git clone https://github.com/natelandau/sessionmemory
cd sessionmemory
uv sync                  # install dependencies
uv run duty lint         # ruff, ty, typos, yamllint, shellcheck, prek
uv run duty test         # pytest with coverage

CLAUDE.md records the conventions this project holds itself to.

License

MIT. See LICENSE.

{
  "name": "sessionmemory",
  "owner": {
    "name": "Nathaniel Landau"
  },
  "description": "Durable memory for coding agents: a CLI and the session hooks that feed it.",
  "plugins": [
    {
      "name": "sessionmemory",
      "source": "./",
      "description": "Project memory: one field of markdown pages per project, captured by an end-of-session sweep and injected at session start."
    }
  ]
}