Local‑first memory for coding agents with Graft
Описание и применение
Graft предоставляет локальное хранилище памяти для кода‑агентов, позволяя сохранять и повторно использовать исправления и решения, найденные во время работы.
Задача
Coding‑agent sessions lose learned fixes and decisions after the session ends, leading to repeated work and loss of knowledge.
Результат
Graft enables coding agents to remember and reuse fixes, constraints, and gotchas locally, reducing redundant work and improving efficiency.
Рабочий процесс
- 1. Run
graft query <question>to check if the issue has been seen before and get a confidence‑gated result. - 2. If useful, run
graft retrieve <question>to get ranked memory nodes that may help. - 3. Use
graft explore <keyword>to navigate related memories via semantic and keyword relationships. - 4. When a new fix or insight is found, insert it with
graft insert --title <title> --body <body> --keyword <kw1> --keyword <kw2> .... - 5. Continue coding; the agent can call Graft via CLI, subprocess, REST, or MCP to access memory during work.
Доказательства
Graft is an open‑source project for keeping that kind of knowledge available locally across later agent work.
The core runtime is local: a CLI communicates with a local daemon, which uses SQLite storage alongside FTS5 and sqlite‑vec.
- Дата записи: 2026-09-20T23:45:40.624Z