Aulafy
ESEN
Courses
Courses/Agents and automation/Shared Persistent Memory
Lesson

Shared Persistent Memory

When multiple agents work over days, chat context isn't enough. You need external memory: task state, decisions, retrievable documents, and a way to forget what's no longer useful.

  • Separate working memory, vector memory, and human wiki.
  • Prevent multiple agents from writing contradictory facts.
  • Design decay, permissions, and traces for shared memory.
TerminalCode
memory_layers:
  hot:
    storage: "task_board"
    content: "estado actual, bloqueos, owner"
  warm:
    storage: "qdrant"
    content: "decisiones, conversaciones resumidas, documentos"
  cold:
    storage: "obsidian"
    content: "wiki revisada, fuentes y reglas estables"
  rules:
    write_requires_source: true
    conflicting_fact: "create_review_task"
    decay_days: 30

What to Store

  • Decisions made and the rationale.
  • Repeated errors and the applied fix.
  • Explicit user preferences.
  • Task state and handoffs.
  • Stable sources with dates.
TerminalCode
memory_event:
  type: "decision"
  scope: "project:aulafy"
  fact: "Las lecciones deben citar fuentes oficiales cuando tratan herramientas vivas."
  source: "editorial_policy"
  confidence: 0.95
  expires_at: null
  writer_agent: "planner"

Official Sources

  • LangGraph Persistence
  • Qdrant documentation
  • Obsidian Flavored Markdown
  • Hermes Agent documentation
Local Multi-Agents, Memory, and LoopsMCP Verification Before Merge