engram
由 dannydvm 提供
DevOps & Cloud简介
Persistent semantic memory layer for AI agents.
描述
该技能的功能
Engram is a local-first persistent memory layer for AI agents, backed by SQLite and LanceDB with Ollama embeddings. It stores typed memories (facts, decisions, preferences, events, relationships) across sessions and surfaces them via semantic search and context-aware recall. Memory decay, deduplication, agent scoping, and graph relations are built in.
Runs entirely local with no token cost and no cloud dependency, so memory persists through context compaction and session crashes without sending data to an external service.
何时使用
- Recalling a client's communication preferences from a session three weeks ago
- Storing a database architecture decision so future agents don't revisit it
- Ingesting a meeting transcript to automatically extract structured memories
- Searching past project milestones before drafting a status update
- Keeping memories isolated per agent so two bots don't share unrelated context
示例工作流
User asks: What did we decide about the deployment setup for the new client?
- 1
Run
engram recall "client deployment setup" --limit 10to surface the most relevant past memories - 2
Review returned memories ranked by semantic similarity, recency, and salience
- 3
Identify the stored decision memory: "Chose Docker Compose over Kubernetes for simplicity on small clients"
- 4
Check related memories via
engram relations <memory-id>to find linked facts or follow-up events - 5
Return the decision with full context to the user
"We decided on Docker Compose over Kubernetes for this client (stored 2026-01-15). Related: client prefers minimal ops overhead, deployment target is a single VPS."
前置要求
- Ollama installed with the nomic-embed-text model pulled
安装
clawhub install dannydvm/engram运行 CLI 命令,或把这段提示词粘贴到 Claude Code、Cursor、Codex 或你在用的任意编程智能体里。