hermes - 💡(How to fix) Fix Harden memory capture and recall bounds for Supermemory/provider-backed memory [1 pull requests]

Summary

Several memory-provider and memory-recall paths can unexpectedly increase token usage or push large payloads into the active agent context. This is most visible with Supermemory-backed setups, but the underlying fixes are general memory hygiene and should apply to any provider-backed memory/search integration.

Problems observed

1. auto_capture: false should fully disable automatic session-end ingestion.

   • Expected: only explicit memory writes should be persisted when auto-capture is disabled.
   • Risk: if session-end capture still runs, memory quota can be consumed very quickly and users lose trust in turning memory on.

2. capture_mode: explicit should remain explicit.

   • Expected: explicit capture mode should not be normalized/coerced into broad capture.
   • Risk: users choose conservative memory behavior but still get broader ingestion than intended.

3. Supermemory capture should safely handle multimodal/list/dict message content.

   • Expected: content cleaning should stringify text-bearing structured content and omit raw image/media payloads.
   • Risk: multimodal gateway turns can raise type errors such as expected string or bytes-like object, got 'list', or store unusable media-shaped payloads.

4. Memory/search/profile tools need output caps at the tool boundary.

   • Expected: provider-backed recall/search/profile results should respect a configurable character cap and append a visible truncation notice.
   • Risk: recall can flood the active context, trigger compression, degrade latency, and make memory feel worse as it grows.

5. Explicit memory writes should include useful source metadata.

   • Expected: explicit saves can include metadata such as source/tool/container tag/session context where appropriate, without logging secrets.
   • Risk: later cleanup, debugging, or auditing is harder than necessary.

Suggested implementation

• Treat memory.auto_capture: false as authoritative in session-end memory ingestion paths.
• Preserve capture_mode: explicit during config/provider normalization.
• Add a Supermemory content-normalization helper that:
  • accepts strings directly,
  • extracts text from OpenAI-style content arrays,
  • stringifies simple dict/list text content,
  • skips or summarizes image/media payloads rather than storing raw payloads.
• Add a shared helper for memory-search result caps, e.g. driven by memory.search_result_char_limit.
  • Apply it to session/memory search and provider-backed profile/search output.
  • Make truncation explicit in the returned text, not silent.
• Add tests covering:
  • auto_capture: false prevents session-end persistence,
  • explicit capture mode is preserved,
  • multimodal/list content does not crash Supermemory cleanup,
  • memory/search/profile output respects the configured cap,
  • explicit writes still succeed and include non-sensitive metadata.

Why this matters

A user can have a paid memory backend and still need conservative capture. Turning off broad automatic capture should not make memory useless; it should make memory curated. Explicit save/search flows still need to work well, while automatic ingestion and recall payloads stay bounded.

This would make Hermes safer for long-running gateway deployments where memory is always available but should only persist durable facts when explicitly requested or when the agent intentionally chooses to save something.