Skip to content

Expose privacy-safe memory retrieval receipts #985

@caioribeiroclw-pixel

Description

@caioribeiroclw-pixel

Supermemory is positioning the right problem very clearly: memory is fragmented across Cursor, Claude, VS Code, Claude Code, Cline, etc., and MCP can make recall portable.

One observability/debugging gap I keep seeing in shared-memory workflows is: after a cross-client recall happens, what privacy-safe evidence can a user or maintainer inspect/share without dumping the raw memory, prompt, transcript, or private repo text?

For example, when someone says “Claude did not remember the PDF I added in Cursor” or “the wrong project memory was retrieved”, the useful artifact is not the memory body. It is a small receipt of what the memory layer returned and whether the client/harness loaded it.

A minimal receipt shape could be:

memory.search.returned
  memory.provider = supermemory
  client.name = cursor | claude-code | vscode | ...
  project.id_hash
  query.hash                 # not raw query by default
  result.count
  result.ids_hash
  result.score_bucket[]
  result.content_hash[]      # hash of returned chunks, no raw memory text
  snapshot.id/hash
  latency_ms

context.input.loaded         # emitted by client/harness if it entered the session
  source.uri = supermemory://...
  delivered.hash
  estimated_chars
  activation = mcp_search | automatic_recall | manual
  duplicate.suppression_policy

That separation matters because the MCP server can prove “these memories were returned”, while the client/harness proves “this context actually entered the model/session”. Both can stay privacy-first.

Why I think this is useful:

  • user bug reports can include receipts without exposing private memories;
  • maintainers can distinguish “search returned nothing” vs “client did not load returned memory”;
  • teams can audit cross-client memory usage without enabling raw prompt/memory logging;
  • it creates a bridge to OpenTelemetry-style agent traces without making Supermemory responsible for the entire agent runtime.

I have a small public fixture exploring this general receipt pattern here, if useful as prior art: https://github.com/caioribeiroclw-pixel/pluribus/tree/main/examples/context-input-evidence

The concrete ask is small: would Supermemory consider an optional debug/export mode that emits privacy-preserving memory search/result receipts, or is this better as an external wrapper around the MCP client/server boundary?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    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