mem0

Learn more · Join Discord · Demo

📄 Building Production-Ready AI Agents with Scalable Long-Term Memory →

---

Learning Project (Personal Fork)

This repository is my learning fork of mem0, focused on improving two concrete memory-system failure modes through iterative evaluation: This repository is an independent learning fork and is not affiliated with mem0ai.

1. Temporal-aware retrieval (recency vs historical query intent)
2. Conflict resolution (replacement vs augmentation)

I treated this as a small research-style engineering project: define failure taxonomy, build controlled datasets, run repeated evaluations, and update architecture based on measured regressions.

What I Changed

• Temporal reranking module

  • Added half-life based temporal score fusion in retrieval.
  • Added query-aware temporal gating for recency-style queries.
  • Added hard historical gating so explicit lookback queries disable freshness bias.
  • Added configurable knobs in MemoryConfig:
    • temporal_half_life_days
    • temporal_weight_override
    • temporal_base_decay_weight

• Conflict resolution module

  • Added append-only conflict metadata:
    • superseded_by, supersedes, source_type
  • Added resolver-side semantic guard against destructive false links.
  • Introduced link_type routing in extraction + resolver:
    • supersede -> eligible to hide old memory
    • augment -> keep both memories visible
    • related -> keep both memories visible
  • Updated conflict eval runner to distinguish extractor observations from resolver behavior.

Evaluation Assets

• evaluation/temporal/controlled_temporal_eval_set.json (40 cases)
  • explicit recency / implicit recency / historical lookback / compact gap / atemporal
• evaluation/temporal/run_controlled_temporal_eval.py
• evaluation/conflict/controlled_conflict_eval_set.json (21 cases)
  • strong conflict / weak conflict / non-conflict / chain update / additive-like non-replacement
• evaluation/conflict/run_controlled_conflict_eval.py
• Detailed iteration log: phase1_experiment_log.md

Key Results

Temporal evaluation (controlled 40-case set)

From evaluation/temporal/controlled_temporal_eval_report.json:

• explicit_recency top1: 0.0 -> 1.0
• implicit_recency top1: 0.0 -> 1.0
• compact_gap top1: 0.0 -> 1.0
• historical_lookback top1 remains 1.0 after historical gating fix
• atemporal_fact atemporal drop rate remains 0.0

Conflict evaluation (controlled 21-case set)

Final run from evaluation/conflict/controlled_conflict_eval_report.json:

• case_pass = 20/21
• assertion_fail = 1
• assertion_observed = 2
• destructive supersession misfires across iterations: 5 -> 3 -> 0

Interpretation:

• Visibility correctness is the primary acceptance target and is now stable.
• Remaining failure is extractor-side noisy linking (related) without incorrect hiding.

Design Rationale

• Append-only memory safety first: old memories are not deleted; they remain auditable via ID/history.
• Extractor vs resolver split: extraction proposes links, resolver enforces visibility policy.
• Policy through link_type: prevents "linked = hidden" coupling and cleanly separates replacement from augmentation.