userId is not threaded into the extraction prompt — first-person facts get unanchored

[amiller]

Summary

When /v1/ingest receives content with first-person language ("I'm working on X", "my favorite Y is Z"), the extractor produces triples whose subjects are bare strings like "User", "I", or "User's private library" — never the userId that was passed in the request.

userId is correctly honored as a storage-side filter (search and recall scope correctly), but it never reaches packages/memory/src/providers/extract.ts or the extraction prompt. The extractor has no knowledge of who "I" refers to.

Reproduction

Standalone Python script: clients/python/examples/repro_extractor_no_user_identity.py (no dep on the in-flight #143 client — uses raw httpx).

Runs against a live selfhost (http://localhost:4010), repeats 5x, ingests one turn under userId="andrew" containing both first-person facts ("my favorite color is teal", "I'm working on matthammer") and a third-party named entity reference (research about "Shashank, founder of Contexto").

Observed (Gemini Flash, MEMORY_EXTRACT_PROVIDER=gemini)

Across multiple 5-run batches:

matthammer subject in extracted triple	Anchored to userId="andrew"?
"User's private library"	❌
"User" (User → is working on → matthammer)	❌
"User"	❌
"User"	❌

0/N runs ever produce a subject containing andrew or any binding to the userId despite it being passed in the request.

Knock-on effect (the bigger issue)

When the extraction context contains a more "concrete" named entity, the extractor can latch onto that entity as the substitute for first-person references — producing triples like (Shashank, prefers, teal) when the user actually said "my favorite color is teal".

Real-world manifestation we hit while wiring Contexto into hermes-agent: a research subagent's findings about a real person mixed with primary-agent self-reflection, and the extractor started attributing the user's own projects to the researched person across hundreds of triples (e.g. matthammer → is a project of → Sha). Hard to repro deterministically in isolation, but the unanchored-subject bug is the root enabler.

Proposed fix (small, drop-in)

Thread userId into the extraction prompt in packages/memory/src/providers/extract.ts. Roughly:

async function callExtract(cfg, model, apiKey, text, userId?: string) {
  const userHint = userId
    ? `\n\nNote: First-person language ("I", "my", "me") in the conversation below refers to the user with id "${userId}". Use that as the subject when extracting first-person facts.\n\n`
    : "";
  const promptText = `${EXTRACT_PROMPT}${userHint}${text}`;
  // ... rest unchanged: pass promptText into the LLM call
}

And plumb userId through:

• extract() in extract.ts
• Memory.ingest() in memory.ts:103
• router.post('/v1/ingest') in router.ts:152 (already has userId in scope from the request body)

Environment

• Selfhost stack built from Dockerfile.selfhost (with a couple of patches — pnpm@9 pin and node:22-slim runtime — happy to send those as a separate PR if useful)
• MEMORY_EXTRACT_PROVIDER=gemini, MEMORY_EMBED_PROVIDER=gemini, gemini-2.5-flash for extraction, gemini-embedding-001 for embeddings
• Found while integrating Contexto into hermes-agent (companion plugin branch). Python client for selfhost in flight at feat: add Python client at clients/python/ #143.

Happy to send the fix as a follow-up PR if the team's open to the approach. The fix would also break ties for the knock-on attribution drift, since the model would have an explicit anchor for first-person rather than reaching for the most-salient nearby named entity.

[amiller]

Fix PR opened: #145

Summary of what's in it:

• Threads userId through /v1/ingest → Memory.add → extractFn → callExtract
• Two changes inside extraction when userId is set: (a) one-line primer prepended to the prompt, (b) source-text speaker label rewritten from "User:" to "${userId}:"
• The text-rewrite is the load-bearing piece; appended-instruction primers alone proved unreliable on gemini-2.5-flash
• Backwards compatible: when userId is not provided, behavior is unchanged
• Includes 8 unit tests for labelSourceText

Re-running the repro from this issue against the patched server: 5/5 runs anchor first-person facts to userId, 0/5 produce bare "User" subjects (was 4/5 before).