Pensieve

A personal, agent-driven memory that survives across sessions.

You talk to an AI assistant (Claude Code) every day, but it forgets everything between sessions. Pensieve is the long-term memory you control: any time in a conversation you can say "add this to pensieve" to capture something, or "what do I know about X?" to pull it back — and it's there, organised, current, and yours.

The metaphor: deliberately draw out a strand of thought and deposit it in the vessel to revisit later.

It's a small Python + SQLite engine with two front doors: an MCP server the agent uses, and a CLI for you. Your memory lives in a single file at ~/.pensieve.

Philosophy

Five ideas shape every decision in Pensieve:

1. An information lake, not a project manager. It stores what you know and lets you recall it. It deliberately has no tasks, statuses, or due dates — the agent infers state by reading, the store just holds the information.

2. Deliberate on both ends. Pensieve never acts on its own. It writes only when you say "remember this," and recalls only when you ask. No silent saves, no auto-loading your memory at the start of a session. You stay in control of what goes in and what comes out.

3. Structure emerges; you don't design it upfront. You don't build a taxonomy. You keep a few top-level streams (the domains you actually work in) and drop notes in. The people, orgs and topics your notes mention become entities automatically, and one that keeps recurring earns its own thread. Organisation is a consequence of use.

4. Notes are the atoms; everything else references them. A note can stand alone or live in several streams at once. Entities and threads are views over notes, not owners of them — so removing a topic never destroys a note that's also about something else.

5. Point at the world, don't copy it. Attach a repo, file or URL as an asset — a by-reference pointer with a one-line "how to use me" hint. Pensieve stores the pointer and reads it on demand; it never crawls your disk or follows a link on its own.

Want the reasoning in depth — the design decisions, the full model, the tradeoffs? docs/philosophy.md.

The model

Recall has three lenses: by name (find), by content (search — full-text, stemmed, ranked), and by time (recent — what changed lately).

Install

Assumes Python 3.12+.

Recommended: from PyPI

Pensieve is a standard MCP server, so the quickest way in is pipx (or uv):

pipx install pensieve-mcp     # or: uv tool install pensieve-mcp

This puts pensieve (the CLI) and pensieve-mcp (the MCP server) on your PATH. Then point your agent at it. For Claude Code:

claude mcp add --scope user pensieve -- pensieve-mcp

For any other MCP client, register the pensieve-mcp command as a stdio server. To run it ad-hoc without installing, uvx pensieve-mcp works too.

Pensieve is listed in the official MCP Registry as io.github.praveen-ilangovan/pensieve, so registry-aware clients can discover and install it directly.

Note: the PyPI package is the engine (CLI + MCP server). The judgment-bearing Claude skill (the capture/fetch flows) ships with the repo, not the wheel — if you want it, use the script install below, which drops it into ~/.claude/skills/pensieve/ for you.

Agent setup from source

git clone git@github.com:praveen-ilangovan/pensieve.git
cd pensieve
./install.sh          # Claude Code (backward-compatible default)
./install.sh codex    # Codex
./install.sh all      # both

install.sh checks prerequisites up front (and changes nothing if any are missing), ensures pipx, installs Pensieve (pensieve + pensieve-mcp on your PATH), installs the selected agent skill, and registers the MCP server in that agent's user configuration. Claude's skill lives in ~/.claude/skills/pensieve/; Codex's user-scoped $pensieve skill is symlinked at ~/.agents/skills/pensieve/, so it is available from every repository. It's idempotent — re-run anytime to pick up updates.

The Codex MCP registration pins PENSIEVE_HOME=~/.pensieve, ensuring Codex reads and writes the same personal store as Claude even when Codex is launched from this repository (whose .env intentionally points development commands at .local/manual). The engine and MCP tools are unchanged; the Codex skill supplies the same deliberate capture/recall discipline as the existing Claude adapter.

Then restart the selected agent and, from any directory:

"what streams do I have? check pensieve"

Your memory lives in ~/.pensieve.

Using it (with the agent)

Pensieve shines through the agent — you speak naturally, it does the judgment and calls the tools. Everything below is just talking to Claude Code.

Set up your domains (do this once, deliberately):

"Create a pensieve stream called Career — for my job search and work."

Capture — any time something worth keeping comes up:

"Add this to pensieve: had a call with Maya about the platform role; she's reviewing my portfolio."

The agent filters for what's durable, routes it to the right stream, and recognises that "Maya" is a person worth tracking — without you managing any of that. If a note spans two domains (say, a talk that's relevant to both your career and a side-project), it files one note in both.

Recall — pick the lens that fits the question:

"What do I know about Maya?" · "What did we decide about salary?" · "Catch me up — what's changed lately?"

Point at live context — so the agent knows where to read:

"Add my project repo at ~/projects/acme as an asset on the side-projects stream — hint: read README.md first." Later: "pull up the acme repo." (It follows the pointer only when you ask.)

Promote — when something recurs, the agent proposes it:

"Maya's come up across 5 notes — want her own thread under Career?"

Remove / restore — everything is soft and reversible:

"Remove that note." / "Actually, bring it back." / "I'm done with the X stream."

Using it (the CLI)

The same engine is a CLI too — handy for a quick check or scripting without the agent (pensieve stream list, pensieve search "…", …). Full command reference with examples: docs/cli.md.

How it works (under the hood)

• SQLite store (~/.pensieve), self-migrating via Alembic on first use.
• Full-text search via SQLite FTS5 + a porter stemmer (so "pricing" recalls "priced").
• A clean ports/adapters core: services depend on a storage port with two interchangeable backends (SQLite + an in-memory double), kept honest by a conformance test.
• The MCP server and CLI are thin "op" layers; the judgment (what to keep, how to resolve an entity, when to promote) lives in the agent skill (adapters/claude/).

Develop on it

make install         # poetry env + pre-commit hooks
make test            # unit + integration
make eval            # deterministic engine evaluators
make check           # ruff (lint+format) + mypy
make manual ARGS="stream list"   # run the CLI against the local dev store

This repo is a self-contained dev environment — it never touches your real ~/.pensieve. The in-repo MCP server and CLI use a local dev store (.env → .local/manual); the global install (./install.sh) is what points at ~/.pensieve.

Design notes live in plans/ (one file per slice, plus plans/roadmap.md) and docs/ (philosophy.md — the why + model, cli.md — commands).

Status

Working and in daily use: streams · threads · notes · entities · promotion · assets · search · recency · multi-stream notes · soft remove/restore — all via CLI and MCP. Built and validated slice by slice with a real agent. See plans/roadmap.md for what's next.