fix(hindsight): frame recalled memories as external evidence

[EndeavorYen]

Summary

• Frame default Hindsight retain context as conversation evidence rather than identity.
• Strengthen the default recall preamble so retrieved memories are not treated as system instructions.
• Add regression coverage for the default retain and recall framing.

Problem / Root Cause

Hindsight auto-retains full conversation transcripts. When a user discusses fiction, roleplay, documents, or code, later recall could present that content without enough provenance, making it easier for the agent to treat external content as user or agent identity.

Fix

The built-in defaults now distinguish stable user/session facts from external content discussed in past conversations, while preserving user-provided retain_context and recall_prompt_preamble overrides.

This is intentionally narrower than a full content-type classifier: it adds a low-risk boundary at both retain and recall time while keeping existing Hindsight behavior and APIs intact.

Addresses #21709.

Limitations / Future Work

This PR reduces identity drift risk by adding provenance framing. It does not classify, drop, or tag fiction/document content at ingestion time; a future follow-up could add explicit content-type metadata if Hindsight exposes the right storage/retrieval hooks.

Related PRs Checked

• fix(hindsight): sync bank_mission and bank_retain_mission to Banks API #18789 syncs bank_mission / bank_retain_mission to the Banks API; it does not change default retain/recall framing.
• fix(hindsight): use descriptive context and stable document_id in sync_turn #8463 improves sync_turn context/document IDs; it does not address fiction/persona contamination on recall.
• fix(memory): guard stale recall and unsafe learning #18372 guards stale recall and unsafe learning; it does not distinguish external narrative content from identity.
• Tag Hindsight retained memories with chat topics #20662 and Replace cumulative Hindsight transcript snapshots #20664 adjust Hindsight topic tags / cumulative transcript snapshots; they do not add persona-boundary framing.

Validation

• .venv/bin/python -m pytest tests/plugins/memory/test_hindsight_provider.py -q
• .venv/bin/python -m py_compile plugins/memory/hindsight/__init__.py

Platforms Tested

• macOS, Python 3.11.15

[teknium1]

Thanks for the focused provenance-boundary mitigation. The reported path is still present on current main: Hindsight retains the full serialized transcript with the generic default context (plugins/memory/hindsight/__init__.py:688, :1621-1673) and injects recalled results under a generic header (:1477-1482). The proposed default framing therefore addresses the direct current behavior.

Problems

• The changed README is not the only user-facing configuration reference. website/docs/user-guide/features/memory-providers.md:432 and website/i18n/zh-Hans/docusaurus-plugin-content-docs/current/user-guide/features/memory-providers.md:362 would still publish the old default after this change.

Suggested changes

• Update those two configuration tables when salvaging the README/runtime change.
• GitHub currently reports this PR mergeable=false; replay the small default substitutions against the current provider rather than relying on the stale line positions.

Automated hermes-sweeper review.