LaPis 💎

🇵🇭 lapis — pencil, to write 🪨 lapis — stone, precious gem 🧠 LaPis — memory for Pi

Persistent memory for the Pi coding agent. One SQLite database, zero cloud dependencies, zero API keys.

PiMemory Architecture

curl -fsSL https://raw.githubusercontent.com/genegulanesjr/PiMemoryExtension/main/install.sh | bash

Restart Pi and memory auto-wires on session start.

• Remembers across sessions — decisions, bugfixes, patterns, discoveries persist
• Auto-injects context — next session starts with relevant memories loaded
• Code indexing — web-tree-sitter (WASM) AST parses JS/TS/Go/Python/Rust/SQL files, searchable by issue description
• Trust scoring — memories linked to changed code lose trust; stable code boosts it
• Deduplication — trigram overlap prevents duplicate saves (≥85% auto-merges, 60–84% warns)
• Workspaces — formal project isolation with create/list/archive
• Session lifecycle — auto-recovery of incomplete sessions, trust recovery on close
• Zero servers — single Node.js CLI + SQLite, called on demand by Pi. Zero Python dependency.
• Dream Cycle — every 10 sessions, cleans superseded, zero-recall, stale corrections, and replaced configs
• Update & delete — update memories in-place instead of spawning correction entries

Unlike compact (housekeeping: vacuum, FTS optimize), the Dream Cycle targets staleness — information that is no longer accurate or useful:

Age alone is NOT a signal. A 6-month-old valid decision stays. A 1-day-old superseded one goes.

Create ~/.pi/memory/config.jsonc to override defaults. The file supports JSONC (JSON with // and /* */ comments).

{
 // Database file path (default: ~/.pi/memory/memory.db)
 "db_path": "~/.pi/memory/memory.db",
 // SQLite WAL autocheckpoint threshold (default: 1000)
 "wal_autocheckpoint": 1000,
 // SQLite busy timeout in milliseconds (default: 5000)
 "busy_timeout_ms": 5000,
 // Search ranking weights (must sum to 1.0)
 "ranking": {
 "fts_relevance": 0.4,
 "recency": 0.3,
 "trust": 0.15,
 "recall": 0.15
 },
 // Deduplication thresholds (0.0 - 1.0)
 "dedup": {
 // Auto-merge duplicates at or above this similarity
 "auto_merge_threshold": 0.85,
 // Flag potential duplicates at or above this similarity
 "warning_threshold": 0.60
 },
 // Run compaction every N sessions for a project (default: 5)
 "compact_every_n_sessions": 5,
 // Path to tier config file (default: ~/.pi/memory/tier.jsonc)
 "tier_config_path": "~/.pi/memory/tier.jsonc"
}

All options are optional — only include the ones you want to change. Missing values use built-in defaults.

• Node.js ≥ 22.5 (built-in node:sqlite) or Node.js with better-sqlite3
• No Python dependency — code parsing uses web-tree-sitter (WASM) in-process
• No API keys or cloud services needed

JavaScript, TypeScript, TSX, Go, Python, Rust, SQL

The wire format (wire-format.js) uses compact encoding (columnar CSV with path interning, uniform-column hoisting, and internal-ID stripping) to reduce the token footprint of analysis responses inside Pi's context window.

Measured via bench/bench-tokens.js — runs real CLI commands against indexed repos, passes output through compactResponse(), and compares byte sizes.

Key findings:

• Dead-code sees the biggest gains (42–47%) — the signals and confidence fields are uniform across rows and get hoisted, while symbol_id is stripped
• Coupling benefits from prefix interning on shared file paths (33–39%)
• Larger repos trend higher — Aether (154 files, 1,359 symbols) edged out PiMemoryExtension (38 files) at 37% vs 36%
• Cycles and very small result sets show no savings (no homogeneous lists to encode)

All transforms are lossless round-trip — verified by 217 tests in test/wire-format.test.js.

Run it yourself:

node bench/bench-tokens.js

MIT