Fenglei Memory · 风雷记忆 · 学而非存
Most AI agents today treat memory like a warehouse: pile up every chat, every file, every context, and hope retrieval finds what matters. It doesn't work. Information grows, but the truly reusable methods get buried deeper every day.
Fenglei Memory takes the opposite approach. It teaches agents to learn from collaboration instead of just storing content, distilling reusable experience into sharper and sharper forms, until stable, callable judgment standards, operating procedures, and skill packs emerge.
Why It Matters
The old pattern usually looks like this. One event happens, then record one entry. One conversation ends, then save one snippet. One fix is done, then drop one log. The result: memory becomes a warehouse, retrieval becomes fishing with a needle. You have done it before, yet you still have to rethink from scratch. Stored a lot, but never really learned. That is not long-term collaboration. That is hoarding.
The Four Layers of Long-Term Memory
Fenglei Memory organizes memory into four layers, and the higher the layer, the closer to learned.
L1 World Facts: stable background such as leader, conventions and project direction, stored in MEMORY.md.
L2 Experience Facts: traces of a real action, what changed and how it was located, stored in daily logs named YYYY-MM-DD.md.
L3 Observations: patterns distilled from repeated facts, stored in MEMORY.md or spec docs.
L4 Skills and Mental Models: stable patterns promoted into callable skills, stored as skill packs, SOPs and checklists.
The goal is to keep pushing facts upward, from raw records to reusable skills.
Core Workflow
Record facts, only the ones with long-term value. Ask what did I learn in each round. Promote on repetition. Skills take priority over raw records.
Promotion triggers, any one of these. Same action repeated two or more times. Same pitfall hit two or more times. Forgetting once causes significant rework. The pattern holds across multiple tasks.
Memory is not written once and permanent. It has a full lifecycle: Write, Retrieve, Reflect, Promote, Clean.
Review cadence: daily scan for promotable items, weekly merge duplicates, monthly formalize stable patterns into skill packs.
What Makes It Different
Zero dependency: it is a methodology, not infrastructure. No vector database, no heavy framework, works in any agent runtime.
Skills over records: the end goal is forming mental models, not archiving.
Inspired by Hindsight: retain, recall, reflect, with the final destination being learned behavior, not a bigger archive.
Get It
Fenglei Memory is open source and free.
https://github.com/Kim-FengLei/fenglei-memory
MIT licensed. Star it, fork it, and give your agents a memory that actually learns.
Built by KIM, Founder of FENGLEI YI, 风雷益, 天施地生，其益无方。
Top comments (0)