agent-memory

原文涵盖的主题

• Agent Memory — promotion is earned, not asserted
• The problem
• The four tiers
• The gates
• Use it
• Hard rules
• Forcing questions
• More in this skill

Portability: stdlib only. No database, no embeddings, no network, no LLM calls. A project's CLAUDE.md is a memory system with one tier and no eviction: every durable fact and every passing preference land in the same always-loaded file, until the important lines are diluted by the incidental ones. Facts learned mid-session vanish at teardown unless someone writes them down. The fix is not more storage — it is a promotion ladder. A claim earns its way toward always-loaded context by recurring; a human confirms the last step. Tiers are distinguished by injection policy, not storage format. | Tier | Holds | Injected | Committed | |---|---|---|---| | L0 | raw session transcripts | never | no (already on disk) | | L1 | candidate atoms | on relevance, at prompt time | no (gitignored) | | L2 | this project's context | every session start | yes, after adopt | | L3 | stable cross-project persona | always | yes, after adopt | Nothing moves up because it sounded important. It moves up because it recurred. L0 → L1 — an explicit marker fires (a directive, a correction, a stated preference, a named lesson, a reproducible failure). Rule-based, high precision, deliberately low recall. L1 → L2 — ≥

这里只展示节选，请以完整原文和当前要求为准。