CodeAbra/iai-personal-memory-engine

A cyber brain for your AI. It never forgets a detail, remembers exactly what you said, and learns how you work over time. Free, local, works with Cursor, Claude Code, Codex, OpenClaw, Hermes and more. MIT.

485 ★54 forksPythonUpdated 1d ago

What you need to know

iai-pme is a fully local, ambient personal memory engine for AI coding assistants — it captures every turn of every session verbatim, organizes captures over time into a personal memory map, and injects a relevant memory slice at the start of each new conversation via MCP-over-stdio; no account, no API key, no telemetry.

Install

pip install iai-pme then iai-mcp crypto init (creates the encryption key)
Claude Code plugin: /plugin marketplace add CodeAbra/iai-personal-memory-engine then /plugin install iai-memory@iai-pme
One-command bootstrap: curl -fsSL https://raw.githubusercontent.com/CodeAbra/iai-personal-memory-engine/main/scripts/bootstrap.sh | bash
Source install: git clone, python3.12 venv + pip install ., cd mcp-wrapper && npm install && npm run build, iai-mcp daemon install, iai-mcp capture-hooks install

Usage

  • Register with Claude Code: claude mcp add iai-pme -- node "$(python -c 'from iai_mcp.cli._capture import _resolve_wrapper_path as w; print(w())')" (or use the plugin)
  • iai-mcp capture-hooks install — wires session-start recall, per-turn recall, and turn/session capture hooks (--target claude|codex|cursor|antigravity|hermes|openclaw|all)
  • iai-mcp doctor runs 27 checks; verify capture via tail ~/.iai-mcp/logs/capture-<date>.log
  • CLI: iai recall / temporal-recall / search / capture / teach (feed a file) / upload / ask / brain (dashboard)

Key features

  • 15 MCP tools including memory_recall (returns hits AND anti-hits that contradict the cue), memory_temporal_recall, memory_recall_structural, memory_search, memory_capture, memory_contradict, memory_reinforce, memory_consolidate
  • Three-tier memory: episodic (verbatim, write-once), semantic (induced summaries), procedural (learned profile knobs), backed by a hyperdimensional substrate
  • Ambient capture/recall hooks — you never say "remember this"; session-start injection bounded to ≤3,000 tokens
  • AES-256-GCM encrypted at rest, key in ~/.iai-mcp/.key; fully local with no telemetry and no network calls beyond the model call your CLI already makes
  • Own Rust engine + storage (Hippo), community-detection algorithm (MOSAIC), and Lilli HD substrate; LongMemEval R@5 0.962; dashboard (iai brain) and Tauri desktop app

Best for

Giving a personal AI coding assistant persistent cross-session memory about the user — fully local, private, with no cloud account or API key.

Caveats

  • Requires Python 3.11 or 3.12, Node.js 18+, ~500 MB disk; macOS and Linux fully supported, Windows is in beta (experimental)
  • Intel macOS has no prebuilt wheel and installs from source (needs Rust toolchain)
  • Losing the encryption key at ~/.iai-mcp/.key loses the memories — back it up
  • Nightly consolidation may make one `claude -p` LLM call through your existing Claude subscription (capped at ≤1% of daily quota)
  • MCP transport is a Unix domain socket (no TCP listener)
Platforms: macOS (Apple Silicon) · Linux (glibc 2.28+) · Windows (beta)Clients: Claude Code · Claude Desktop · Codex CLI · Gemini CLI · Cursor · Cline · Continue.dev · Zed · Cherry Studio · Goose · Aider · any MCP-over-stdio host
Documentation ↗

Reviewed 2026-08-07

Topics

agent-memoryai-memoryclaudeclaude-codecodexcursorencryptedepisodic-memorygeminillm-memorylocal-firstlong-term-memorymcpmcp-memorymcp-servermodel-context-protocolpersistent-memoryrag-alternativesqlitevector-database
View on GitHub ↗
Stars
485★
Forks
54
Language
Python
License
MIT
Created
2026-05-06
Last push
2026-08-05