Skip to content

[FEATURE] Retrieval pipeline in code doesn’t match the paper #65

Description

@BaoBaoGitHub

Feature Request

In the current codebase, retrieval is implemented as keyword/vector/hybrid/rrf/agentic in MemoryManager.retrieve_mem, but I don’t see the MemScene selection step described in the paper (score MemScenes by max MemCell relevance, select top scenes, then pool Episodes and rerank). Foresight validity filtering also seems only partially applied depending on the path. The paper-like flow appears closer to the evaluation scripts than the main service.

Is this because the production pipeline prioritized a simpler, general retrieval flow while the paper pipeline was only implemented in eval code? Are there plans to bring the MemScene-first retrieval algorithm (and consistent Foresight validity filtering across keyword/hybrid) into the main service?


Use Case

What problem does this feature solve?
Ensure the production retrieval pipeline matches the paper (MemScene selection → Episode pooling/rerank → Foresight validity filtering), so reported results align with actual system behavior.

Who would benefit from this feature?

  • End users (application developers using EverMemOS)
  • Contributors/developers
  • System administrators
  • Other: _________________

Impact

How important is this feature to you?

  • Critical - blocking my use of EverMemOS
  • High - would significantly improve my workflow
  • Medium - nice to have
  • Low - minor enhancement

Would this be a breaking change?

  • Yes, this would break existing functionality
  • No, this is backward-compatible
  • Not sure

Implementation

Are you willing to contribute to this feature?

  • Yes, I can submit a pull request
  • Yes, I can help with testing
  • Yes, I can help with documentation
  • No, but I'm happy to provide feedback

Estimated complexity:
(If you have development experience)

  • Small - minor addition or change
  • Medium - requires some refactoring or new components
  • Large - significant architectural changes
  • Not sure

Checklist

Before submitting, please check:

  • I have searched existing issues and feature requests to avoid duplicates
  • I have provided a clear description of the proposed feature
  • I have explained the use case and problem being solved
  • I have considered potential alternatives

Activity

  1. BaoBaoGitHub commented on Feb 3, 2026

    @BaoBaoGitHub
    Author

    Another question: why are foresight and event_log supported only in assistant mode, but not in group mode?

    In fact, in group mode there may be multiple users, and each user would also have their own foresight and event_log.

  2. Linzwcs commented on Apr 7, 2026

    @Linzwcs

    same question

  3. Mikivishy commented on Apr 7, 2026

    @Mikivishy

    same question

  4. cyfyifanchen commented on Jun 6, 2026

    @cyfyifanchen
    Collaborator

    Closing as part of the EverOS 1.0 issue triage. This issue targets the old benchmark/evaluation code path. The current repo uses the 1.0 server API for LoCoMo reproduction; see docs/locomo_benchmark.md. PR #258 adds migration notes for legacy benchmark reports: #258. If this still reproduces with current main and the 1.0 benchmark commands, please open a fresh issue with the exact command and output.

  5. cyfyifanchen commented on Jun 6, 2026

    @cyfyifanchen
    Collaborator

    Reopening after a stricter post-triage audit. This issue is about paper-vs-code retrieval behavior rather than only a retired API or infrastructure stack. It should remain open until maintainers either document the EverOS 1.0 retrieval scope clearly or decide the paper-matching pipeline is out of scope.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions