EverMemOS Adaptation

EverMemOS Adaptation for Game-Dev Memory

Date: 2026-02-16 Source: https://github.com/EverMind-AI/EverMemOS

Why this matters

EverMemOS emphasizes structured long-term memory instead of storing only raw chat text. That matches our goal: memory must stay useful across coding sessions, team members, and toolchains.

Key ideas we are borrowing

• Memory derivation pipeline:
  • Convert one source memory into multiple specialized memories.
  • Keep provenance links so derived memories remain auditable.
• Future-oriented memory lane:
  • Distinguish "what happened" from "what should happen next".
  • Treat deadlines and planned actions as first-class retrieval targets.
• Event-style atomicization:
  • Break noisy notes into concise factual units for better retrieval precision.

Implemented in this repo

API endpoints

• POST /api/memories/:id/derive
  • Derives event_log and foresight memories from a parent memory.
  • Supports dry_run, max_event_logs, max_foresight, and selective toggles.
  • Writes entity_links with relation derived_from.
• GET /api/memories/foresight/active
  • Returns time-aware foresight memories sorted by nearest due date.
  • Supports filtering (project_id, q, include_past, within_days, lifecycle controls).

Core derivation module

• api/src/core/memoryDerivation.ts
  • Sentence candidate extraction from free-form text.
  • Event/future cue detection.
  • Date parsing (absolute + relative cues).
  • Confidence estimation and de-duplication.

How this improves retrieval quality

• Lower noise in recall: event logs are short and factual.
• Better planning support: foresight lane makes upcoming tasks/querying explicit.
• Better explainability: each derived memory keeps evidence + parent linkage.

Next evolution steps

• Add MCP tools for memories.derive and memories.foresight_active.
• Add CLI commands that wrap these endpoints.
• Add evaluator metrics:
  • precision@k for bug-fix recall,
  • deadline hit-rate for foresight effectiveness,
  • latency/cost per retrieval mode.
• Add nightly consolidation job to merge near-duplicate derived items.

Guardrails

• Derived memories are additive; parent memory remains source-of-truth.
• We do not derive from already-derived categories (event_log, foresight) to avoid cascading drift.
• Lifecycle flags (active/superseded/quarantined) remain the primary retrieval safety control.