@rohirik/pi-ltm

OpenLTM long-term memory extension for the Pi coding agent. Shares the local memory database with Claude Code and OpenCode.

Install

pi install npm:@rohirik/pi-ltm

If Pi is already running, use /reload to load the extension. Pi records the package in ~/.pi/agent/settings.json; use pi install --local npm:@rohirik/pi-ltm for a project-scoped install.

Find the extension in the Pi package catalog. The npm package declares the pi-package discovery keyword and a pi.extensions manifest pointing to its compiled entrypoint.

Requirements

Install Bun and make it available on PATH. Pi loads the extension under Node; the extension runs OpenLTM's core MCP server in a Bun child process, avoiding direct Node imports of bun:sqlite. The core package is installed as a dependency.

What it does

• Registers the memory tools returned by OpenLTM's core MCP server, including recall, learn, forget, project context, and memory relationships.
• On Pi's before_agent_start event, waits for the core connection and appends relevant project memories as a ## Prior Knowledge (LTM) block.
• Uses a local SQLite database shared with the other OpenLTM adapters. If Bun or the core server cannot be found, the extension currently skips registration.

The current Pi adapter does not register a compaction hook.

Shared memory

When the Claude Code plugin is installed, Pi uses its database, so both agents share memories:

~/.claude/plugins/data/openltm-OpenLtm/openltm.db

(An older install's OpenLtm-openltm/openltm.db is used if that is the one that exists.)

Override it before starting Pi:

export LTM_DB_PATH="/custom/path/openltm.db"

A memory learned in Pi is available to another OpenLTM adapter using the same database.

Semantic recall with llama.cpp

Recall works out of the box with SQLite full-text search. For semantic recall, run a local llama.cpp embedding server; OpenLTM uses it by default when it is reachable, and no API key is needed:

llama-server -m bge-m3.gguf --embeddings --pooling mean --port 8080

Defaults are LTM_LLAMA_CPP_URL=http://127.0.0.1:8080 and LTM_EMBED_MODEL=bge-m3; export either before starting Pi to change it. If the server is not running, recall stays on full-text search and memory text is not sent anywhere. Use the same model as the other adapters that share the database, since vectors from different models are not comparable.

Recalled memories are wrapped in a <recalled-memory> block that marks them as context from earlier sessions, not instructions.

Privacy and license

Memory is stored locally. Optional embedding or janitor providers can send memory content to the provider you configure; see the repository's security notes. MIT licensed.