Skip to content

Update default prompts to suggest rich metadata (model, agent, task, topic, etc.) #73

@jack-arturo

Description

@jack-arturo

Problem

The current default prompts suggest tagging memories with project name and date, but don't mention metadata at all. This means agents store memories with almost no provenance — you can't tell which agent created a memory, what model was used, what task it was part of, or what the conversation was about.

In single-agent setups this is fine. In multi-agent setups (multiple agents writing to the same memory graph), this becomes a real problem for auditing, debugging, and trust.

Proposed Change

Update any default prompts, system prompt snippets, or documentation that describes memory structure to also recommend a standard metadata block on store_memory calls.

Suggested metadata fields (cheap to store, high value):

Field Example Notes
created_by "hub-developer" Agent name/ID that stored the memory
model "claude-sonnet-4.6" Model used at time of storage
task "PR review automation" Task or job the agent was performing
topic "AutoHub architecture" Conversation or research topic
session_id "agent-abc123" Task/session ID for traceability
source "slack" / "whatsapp" / "code" Where the memory originated
confidence 0.85 Already supported — make sure it's in the prompt

Metadata is cheap. The cost of over-storing provenance is negligible vs. the value of being able to audit memory origins later.

Why metadata over tags?

Tags are hard filters — adding hub-developer as a tag means queries filtering on other tags will miss those memories. Metadata is additive and doesn't affect semantic search or tag-based filtering. Provenance belongs in metadata, not tags.

Acceptance Criteria

  • Default store_memory prompt examples include a metadata block
  • Documentation / README updated with recommended metadata fields
  • Consolidation note: when memories merge, source_agents should be a union array (not overwritten) — flag this as a known limitation until automem consolidation supports it
  • Any example snippets or starter configs updated

Related

  • verygoodplugins/automem: Return metadata fields in recall responses by default
  • verygoodplugins/automem: End-to-end metadata audit

Metadata

Metadata

Assignees

No one assigned

    Labels

    documentationImprovements or additions to documentationenhancementNew feature or request

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions