Pi package + native memory engine for the Pi coding agent. Install Pi first, then install Pi-Memory.

Pi-Memory gives Pi agents a durable local memory layer built from three small pieces:

• a native C CLI for persistence and queries
• a Pi extension for compaction/session lifecycle automation
• a MEMORY.md bridge for human-readable project continuity

Repository: https://github.com/SiliconState/Pi-Memory

Pi session JSONL files are excellent for forensics, but they are not optimized for fast recall. Pi-Memory adds a second layer:

• raw session JSONL for full history
• structured memory for durable decisions, findings, lessons, entities, and project state

That makes long-running work easier to resume after compaction, handoff, or reboot.

Pi-Memory has native source support for:

• macOS
• Linux
• Windows

If a matching prebuilt is absent, scripts/setup.mjs automatically falls back to compiling from source.

pi install git:github.com/SiliconState/Pi-Memory

This is the primary install path because Pi reads the package metadata and installs:

• the extension
• the skill
• the prompts
• the native binary via postinstall

Verify:

~/.pi/memory/pi-memory --version

On Windows:

$env:USERPROFILE\.pi\memory\pi-memory.exe --version

The package structure is ready for npm and bun installs from git or a packed tarball.

Examples:

npm install github:SiliconState/Pi-Memory
bun add github:SiliconState/Pi-Memory

Notes:

• postinstall runs scripts/setup.mjs
• prebuilt binaries are used when present
• otherwise the package compiles from source
• Pi users should still prefer pi install git:... so Pi registers the extension/skill/prompt metadata
• npm publish is not live yet

curl -fsSL https://raw.githubusercontent.com/SiliconState/Pi-Memory/main/scripts/install.sh | bash

This helper is Unix-only.

If pi-memory is not on your PATH, use one of these:

• macOS / Linux: ~/.pi/memory/pi-memory
• Windows: %USERPROFILE%\.pi\memory\pi-memory.exe

BIN="${PI_MEMORY_BIN:-$HOME/.pi/memory/pi-memory}"

# initialize project memory markers
"$BIN" init

# write memory
"$BIN" log decision "Use SQLite for agent memory" \
  --choice "Single local database" \
  --rationale "Simple, durable, low operational overhead"

"$BIN" log finding "Extension syncs MEMORY.md before compaction" \
  --category architecture \
  --confidence verified

"$BIN" log lesson "Wrong project attribution caused noisy sync output" \
  --fix "Pass --project explicitly when needed"

"$BIN" log entity "SessionManager" --type concept --description "Pi session tree manager"

# inspect memory
"$BIN" query --limit 20
"$BIN" search "compaction"
"$BIN" state <project>

# project bridge
"$BIN" sync MEMORY.md --limit 15

For PowerShell, use the Windows binary path directly or set PI_MEMORY_BIN.

Primary record types:

• decisions — what was chosen and why
• findings — facts, discoveries, compaction summaries
• lessons — failures and fixes
• entities — named tools, services, concepts, contracts
• sessions — ingested Pi session metadata and stats
• project_state — current phase, summary, next actions, rollups

Session ingest and shutdown automation also carry:

• session_id links across records
• token/cost rollups
• compaction summaries
• auto-extracted decisions/lessons/entities from session JSONL

Think of it as:

• JSONL = camera footage
• Pi-Memory = engineering notebook

You want both.

The extension automates continuity around compaction and shutdown:

• syncs MEMORY.md before compaction
• stores compaction summaries as findings
• captures the latest user intent for auto-resume after auto-compactions
• ingests the current session JSONL on shutdown
• updates project_state
• syncs MEMORY.md again at shutdown
• performs provider/model failover for compaction-related resilience

Inside Pi you can also use:

/compact-threshold
/compact-threshold 75%
/compact-threshold reset
/compact

npm run setup
npm run doctor
npm run test:smoke
npm pack --dry-run

macOS / Linux

cd native && make

Windows (MSVC Developer Command Prompt)

cd native
nmake /F Makefile.win

• Install Guide
• Architecture
• Extension Lifecycle
• Agent Usage Guide
• Contributing
• Security Policy

MIT