Open source context infrastructure

Open Source Context Infrastructure for AI Agents

Name: Agent Memory System
Author: Gaurav Singh

Agent Memory System gives every repository in a workspace a durable, secure memory layer so project context, execution history, and handoffs survive across folders, repos, Antigravity, Codex, Claude, Cursor, and future agents.

Install Star on GitHub

agent-memory

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

Generated memory/context-index.json

Generated memory/09-agent-guidelines.md

Generated memory/10-agent-worklog.md

$ agent-memory worklog handoff \

--agent codex \

--message "tests pass; next publish pages"

memory/agent-handoff.md is ready

for the next agent.

Owned by RAVBYTE TECHNOLOGIES PRIVATE LIMITED

Company maintained. Community open.

Agent Memory System is maintained by RAVBYTE TECHNOLOGIES PRIVATE LIMITED and founded by Gaurav Singh. The project is MIT licensed, public, and open for issues, pull requests, ecosystem detectors, documentation, and agent integrations.

10+memory files

JSONLagent worklog

CIfreshness gates

PRspublic contribution

Repository memory

One command maps every repo in a workspace.

The scanner handles projects with many folders and repos in one workspace. It detects manifests, routes, APIs, configs, tests, storage hints, documentation, and generated directories, then writes Markdown plus a topic index any agent can read.

Automatic upkeep

Memory changes with the repo.

`agent-memory maintain --since main` detects structural Git changes, refreshes `memory/`, and validates the result so stale context does not quietly drift into reviews.

Agent native

Designed for Codex, Claude, Cursor, and more.

A portable skill wrapper tells agents exactly what to read, when to refresh memory, and how to hand work to the next assistant without losing state.

Graph Intelligence

Understand the blast radius.

Built-in static analysis maps dependencies and architectural layers. Agents can query the graph to find out exactly what files will break if they change an API, eliminating the need to guess.

O(1)Graph Query

0Missing Imports

Proven Results

Scientifically benchmarked.

Our open-source benchmark suite measures exact improvements across 42 rigorous tasks for agents like Antigravity, Codex, and Copilot. Agent Memory System fundamentally changes how autonomous agents navigate code.

+34%Concept Accuracy. Agents solve complex tasks with 100% architectural accuracy.

-45%Fewer Files Traversed. Agents navigate directly to correct files instead of hunting.

-90%Tokens on Recovery. Resume cross-session tasks instantly without re-reading the codebase.

ZeroRedundant Work. Handoffs guarantee Agent B knows exactly what Agent A did.

Security first

Useful memory without leaking secrets.

Memory files should help agents understand the system, not expose credentials. Agent Memory System records environment variable names, blocks obvious secret patterns, avoids generated/vendor paths, and labels inferred content so agents know what needs verification.

No secret valuesEnvironment names only, never `.env` values.

Secret pattern checksJWT-like strings, long hex values, and password assignments are flagged.

Generated path guardrails`node_modules`, `dist`, `.next`, `.venv`, and caches stay out of source ownership.

CI memory checksStructural changes must include refreshed memory in pull requests.

Cross-agent continuity

Start anywhere. Continue anywhere.

Agent Memory System records checkpoints, commands, files, blockers, and next steps in `memory/agent-worklog.jsonl`, then generates `memory/agent-handoff.md` so the next assistant can recover the working state immediately.

Antigravity

Codex

Claude

shared context layer context-index.json + agent-handoff.md + agent-worklog.jsonl

Install

Drop it into any repository.

npx @ravbyte/agent-memory-system@latest init
npm install -g @ravbyte/agent-memory-system@latest
agent-memory scan --json
agent-memory maintain --since main
agent-memory worklog checkpoint --agent codex --message "implemented scanner"
agent-memory worklog handoff --agent codex --message "ready for review"

Open source

Built by RAVBYTE. Open to everyone.

Contributions are welcome: ecosystem scanners, validators, docs, examples, tests, agent skills, and CI integrations. Direct pushes to `main` should be blocked by repository rules; changes should land through pull requests with CI passing.

Contributor path

Fork the repository.
Create a feature branch.
Run typecheck, tests, build, and memory check.
Open a pull request.

Contribute on GitHub