Coding agents that remember why.
Open-source tools for building software with AI agents that keep the project's decisions, causes, and history instead of relearning them every session.
Session 1 — Claude Code
✓ found the bug ✓ fixed it ✓ explained why it had to be that way
A month later — new session, maybe a different agent
"I have no memory of this project."
The same moment, with Byori
"I remember. This bug came from decision #42, which superseded #17.
Option B was rejected after incident #18."
A longer context window does not fix this. What gets lost is not text — it is the chain that explains the code:
incident ──caused_by──> bug ──fixed_by──> decision ──affects──> module
└──supersedes──> previous decision
Vector search recalls a similar paragraph. A graph recalls the reason.
| Project | What it is | |
|---|---|---|
| 🕸 | Byori | A project-first native multi-agent coding workspace. Run Claude Code and Codex side by side over a shared graph memory. |
| 🗄 | ByoriDB | The graph engine beneath it — a semantic graph database in Rust with nGQL, rule inference, bitemporal history, and provenance. |
| 🖥 | ByoriDB Studio | A desktop manager for browsing, querying, and administering a ByoriDB instance. |
Claude Code · Codex
│ MCP + skills
▼
Byori multi-agent coding workspace
│ pinned HTTP/nGQL contract
▼
ByoriDB semantic graph database
│
▼
ByoriDB Studio visual management
Dependencies flow in one direction only. ByoriDB knows nothing about Byori; Byori installs and manages a validated engine release.
curl -fsSL https://github.com/byoridb/byori/releases/latest/download/install.sh | bashThis installs a local single-node ByoriDB, registers the MCP server with the coding CLIs it detects, and installs the memory skill. Everything stays on your machine — the data files and the MCP process are local.
Warning
These are early projects under active development. Do not use them as the only copy of important data.
Apache-2.0 across the ecosystem.