agi-memory

Your AI coding assistant forgets everything between sessions. This remembers.

Decisions you already made, bugs you already fixed, what happened last session, how the codebase fits together — kept in a file on your machine and handed back to the assistant next time, so you stop re-explaining your own project.

Works across Claude Code, Cursor, Windsurf, OpenAI Codex, OpenCode, Antigravity CLI, Aider, Goose, Cline, Roo Code, Crush, Pi and Hermes Agent — one memory, whichever tool you open.

No dependencies, no vector database, no background daemon. ~32MB of RAM, sub-millisecond lookups, works offline. Comparable tools install ~500MB of machine-learning libraries and take 200–500ms per lookup.

The Problem in 10 Seconds

Without agi-memory:

You (Tuesday in Claude Code): "Don't use asyncio.gather here, it corrupts the SQLite write lock."
Claude: "Understood, using sequential writes."
— Next day, fresh session in Cursor —
Cursor: "Let's optimize performance by running this with asyncio.gather!" 🤦

With agi-memory:

— Next day, fresh session in Cursor —
Cursor: "Recalled architecture decision [#14752]: using sequential writes to prevent multi-agent SQLite lock contention." 🎯

Proven by the Numbers

• ⚡ Speed & Footprint: <1ms latency across all 4 layers with ~34MB RAM and 0 background daemons (vs. 1.2GB+ for vector RAG).
• 💰 96% Context Token Savings: Injects a compact ~120-token briefing of recent sessions and active invariants instead of dumping 3,000+ tokens of raw logs every session.
• 🛡️ 100% Dead-End Avoidance: Scored 6/6 Actionability and 2/2 Session Recaps on tests/eval_usage.py — warns the agent if previous work was abandoned or blocked so it never repeats failed attempts.

Why agi-memory? The 4 Cognitive Memory Pillars

Most AI memory architectures solve only a fragment of developer memory while incurring heavy dependencies or requiring background Node.js daemons. agi-memory unifies all four cognitive memory pillars in pure Python stdlib + SQLite (<35MB RAM, <1ms speed, zero external pip dependencies):

Every pillar is scored by its own eval suite — see Benchmarks for measured comparisons against Mem0, Zep, Cognee, LangChain and claude-mem, including a real 13,988-observation production dataset.

Install

# One-line installer (recommended)
curl -fsSL https://raw.githubusercontent.com/kdbhalala/agi-memory/main/install.sh | bash

# Or: Homebrew / PyPI
brew tap kdbhalala/agi-memory https://github.com/kdbhalala/agi-memory && brew install agi-memory
pipx install agi-memory

Then wire up your assistants and initialize a project:

agi-integrate install all     # configure every detected assistant + lifecycle hooks
agi-integrate status          # confirm what was detected and configured
cd your-project && agi-integrate init .

init wires the project and installs an /agi-init slash command in each assistant's own format. Run /agi-init inside your assistant and it reads the codebase and writes the project's rules/ and context/ files.

Full options, including uvx and from-source: Installation.

Documentation

License

MIT