Agent Memory System

Persistent project memory, agent worklogs, and cross-agent handoffs for AI coding tools.

What It Does

Agent Memory System gives any repository a durable memory layer that AI agents can read before they work. It scans the codebase, generates structured Markdown memory, creates a machine-readable topic index, tracks structural changes, and records handoffs so context survives when work moves between Antigravity, Codex, Claude, Cursor, or another assistant.

npx @ravbyte/agent-memory-system@latest init

The project is owned and maintained by RAVBYTE TECHNOLOGIES PRIVATE LIMITED, but it is completely open source under the MIT License and open for public contribution.

🧠 Codeflow Graph Intelligence

AMS includes a built-in headless static analysis engine that tracks symbol-level dependencies, architectural layers, and call relationships.

Instead of your agent guessing what files to change, it can query the graph:

• Blast Radius: Detect exactly what downstream files break when you change an interface.
• Health Scoring: Identifies technical debt, circular dependencies, and high-risk 'God Objects'.
• Architecture Flow: Automatically categorizes files into ui, services, data, utils, etc.
• Breaking Changes: Diffs graph snapshots to detect removed API exports before CI fails.

Run agent-memory graph build to generate memory/repository-graph.json and agent-readable summaries.

📊 Scientific Benchmarks

We maintain a rigorous Benchmark Suite that measures the exact impact of AMS on AI agent workflows.

Across 42 canonical tasks (Understanding, Refactoring, Debugging, Recovery, Multi-Agent), tests show that AMS significantly improves autonomous agent capabilities:

• +34% Concept Accuracy: Agents solve complex tasks with 100% architectural accuracy.
• 45% Fewer Files Traversed: Agents navigate directly to correct files instead of hunting.
• Near-Zero Redundant Work: Cross-session recovery and handoffs resume instantly without needing to re-read the codebase.

The benchmark suite is open source—you can run it on your own repos to measure your agent's performance.

Ownership

• Founder: Gaurav Singh
• Company: RAVBYTE TECHNOLOGIES PRIVATE LIMITED
• Website: https://www.ravbyte.com
• Founder X: https://x.com/gauravchadhry
• Founder LinkedIn: https://www.linkedin.com/in/gauravchadhry/
• Repository: https://github.com/RavByte-AI/agent-memory-system

Why It Exists

AI coding agents are powerful, but they forget the project between sessions and tools. A task can start in Antigravity, continue in Codex, get reviewed in Claude, and lose the working state at every switch.

Agent Memory System fixes that by keeping:

• Repository structure and architecture notes
• API, storage, security, testing, and workflow context
• Agent execution checkpoints
• Commands run and files touched
• Human-readable handoff summaries
• CI-enforced memory freshness checks

Generated Memory

memory/
  00-project-overview.md
  01-repository-map.md
  02-system-architecture.md
  03-development-workflow.md
  04-api-and-interfaces.md
  05-data-and-storage.md
  06-security-and-config.md
  07-testing-and-quality.md
  08-known-issues-and-tech-debt.md
  09-agent-guidelines.md
  10-agent-worklog.md
  agent-handoff.md
  agent-worklog.jsonl
  context-index.json
  README.md

Install And Use

Run once in a repository:

npx @ravbyte/agent-memory-system@latest init

Install globally if you prefer a persistent CLI:

npm install -g @ravbyte/agent-memory-system@latest
agent-memory init

Refresh memory after structural changes:

agent-memory maintain --since main

Check memory freshness in CI:

agent-memory maintain --since main --check

Record agent progress:

agent-memory worklog checkpoint \
  --agent codex \
  --message "implemented scanner validation" \
  --files src/scanner/scan.ts,tests/scanner.test.ts \
  --commands "npm test"

Create a handoff before switching agents:

agent-memory worklog handoff \
  --agent codex \
  --message "tests pass; README needs review" \
  --next "review docs and publish GitHub Pages"

Security Features

• Documents environment variable names, never values.
• Validates generated memory for obvious secret patterns.
• Ignores generated and vendor paths such as node_modules/, .git/, dist/, build/, .next/, .venv/, __pycache__/, and target/.
• Labels uncertain sections as [INFERRED], [INCOMPLETE], or [PLANNED].
• Supports CI checks so stale memory cannot silently pass review.
• Encourages branch protection so all changes go through pull requests and CI.

Open Source Contribution

Public contributions are welcome. Good first contributions include:

• New ecosystem detectors
• Better framework and route inference
• More validators for memory quality
• Improved examples and fixtures
• Agent skill integrations
• Documentation and website improvements

Before opening a pull request:

npm install
npm run typecheck
npm test
npm run build
npm run memory:check

Changes to main should go through pull requests with the Required CI status check passing.

GitHub Pages

The static website lives in docs/ and deploys through GitHub Actions.

Repository

https://github.com/RavByte-AI/agent-memory-system

License

MIT