Memory model - mcp-prose-memory

The memory file is a JSON document with a version, update timestamp, and sectioned arrays of facts. Facts can be plain strings or compact structured facts.

{
  "version": 5,
  "updated": "2025-01-15T10:30:00.000Z",
  "sections": {
    "work": ["Fact 1", "Fact 2"],
    "personal": ["Lives in Berlin"],
    "top_of_mind": [],
    "history": ["Completed project X"],
    "instructions": ["Be concise"],
    "user_preferences": [
      {
        "key": "answer_style",
        "value": "Prefers concise answers",
        "confidence": "high",
        "source": "user_explicit",
        "createdAt": "2025-01-15T10:30:00.000Z",
        "updatedAt": "2025-01-15T10:30:00.000Z"
      }
    ]
  }
}

Sections

Section	Purpose
`work`	Professional context, projects, colleagues, tools
`personal`	Location, preferences, interests, personal facts
`top_of_mind`	Current focus and active tasks
`history`	Past events and completed work
`instructions`	Standing rules and behavioral preferences
`user_profile`	Stable user profile facts
`user_preferences`	Durable preferences
`eyra_project`	Eyra-specific product and architecture facts
`devices_environment`	Durable local environment facts
`workflows`	Repeated workflow preferences
`writing_style`	Writing and tone preferences
`long_term_tasks`	Durable task context
`do_not_forget`	Explicitly requested durable reminders

Limits

Limit	Value
Facts per section	30
Characters per fact	300
Characters per structured key	80

Normalization

Older or partial documents are normalized when loaded:

Missing known sections are added as empty arrays
Facts are trimmed
Structured facts are normalized to compact keys and values
A missing updated timestamp is replaced with the current time
Unknown sections are ignored by the active schema

The schema is intentionally narrow so context stays predictable across clients and restarts.