@tinysquid/pi-memory

DEPRECATED - No longer maintained; use other agent-memory approaches.

Persistent project memory for the pi coding agent

Give your agent a project brain. MEMORY.md lives at the project root and is injected into the system prompt on every session start, and the /remember command distills the current session back into it — decisions, preferences, and lessons that survive across sessions, projects, and model resets.

How it works

┌───────────────┐   session start   ┌─────────────────┐
│   MEMORY.md   │ ────────────────▶ │ system prompt   │
│ (project root)│                   └─────────────────┘
│               │   /remember       ┌─────────────────┐
│  Decisions    │ ◀──────────────── │ current session │
│  Preferences  │  (smart merge)    └─────────────────┘
│  Lessons      │
└───────────────┘

• Session start — walks up from the working directory to the project root (the first ancestor containing .git, AGENTS.md, package.json, Cargo.toml, go.mod, pyproject.toml, pom.xml, build.gradle, Gemfile, or CMakeLists.txt), reads MEMORY.md, and injects the full content into the system prompt.
• /remember — feeds the current conversation (user/assistant messages plus any compaction summary) to the active model, extracts new decisions, preferences, and lessons, and merges them into MEMORY.md. The in-session memory is updated immediately, so the next turn already benefits.

Install

pi install npm:@tinysquid/pi-memory

For local development:

ln -s $(pwd)/memory/memory.ts ~/.pi/agent/extensions/memory.ts

Usage

1. Install the extension and start a coding session in your project.

2. Work with the agent — MEMORY.md (if present) is silently part of its context.

3. When you're happy with how things went, run:

   /remember

4. Review the diff, commit MEMORY.md alongside your code, and every future session starts with that knowledge.

[!TIP] Commit MEMORY.md to your repository so the whole team — and every CI agent — shares the same project memory.

Memory file

MEMORY.md lives at the project root and always keeps the same three sections, one terse line per entry:

## Decisions

- Use pnpm workspaces, not npm
- Vitest over Jest for new tests

## Preferences

- Concise answers, no code comments unless asked

## Lessons

- pi extensions load .ts directly, no build step

[!NOTE] Extracted entries are written in an ultra-terse "caveman" style on purpose: articles and filler are dropped, technical terms stay exact. Less noise in the prompt, more room for substance.

Smart merge

/remember never blindly appends. Incoming entries go through a dedup pass before hitting the file:

• An entry that duplicates an existing one — substring match in either direction, or more than 70% shared words — is dropped.
• An entry prefixed with [update] replaces the existing entry it matches instead of adding a new line.
• Entries marked _none_ by the model are ignored.

Behavior notes

• Memory size is capped at 500 lines. Going over triggers a model-driven compression pass; if compression fails, the uncompressed file is written anyway, and if it's still over the limit a warning asks for manual pruning.
• /remember needs a project root, a selected model, and a non-empty conversation — otherwise it notifies and does nothing.
• Extraction, merging, and compression all run with the session's active model; the LLM call honors ctx.signal, so aborting the session cancels it.
• Missing MEMORY.md is fine — the first /remember creates it.