Current

Memory organization and evidence notes

Benchmark first

Corpora: LongMemEval · QASPER · LongBench

Protocol: Replay sessions in timestamp order without gold-answer fields. Report information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention separately; add evidence Recall@k, write amplification, stored bytes, reader tokens, and latency. Compare raw-history, retrieval-only, extracted-memory, and consolidated-memory baselines under identical context budgets.

These functions link related notes, rank memories for recall, and decide whether retrieved evidence can support an answer.

How it works

Note evolution applies a link threshold and a stricter metadata-evolution threshold to caller-supplied similarities. Salience exponentially decays recency, min-max normalizes recency, importance, and relevance over the candidate set, then returns every weighted contribution. Evidence-note decisions validate per-document relevance and answer support before choosing retrieved evidence, explicitly allowed parametric knowledge, or unknown.

Papers

A-MEM: dynamic note evolution Generative Agents: recency, importance, and relevance Chain-of-Note: sequential evidence decisions

memory_evidence.py · current

from mari_components.knowledge import (
    MemorySignal, plan_note_evolution, rank_salient_memories,
)
from mari_components.verification import (
    EvidenceNote, decide_from_evidence_notes,
)

evolution = plan_note_evolution(new_note.id, similarity_by_note_id,
    link_threshold=0.72, evolution_threshold=0.91)
salient = rank_salient_memories([
    MemorySignal(memory_id=m.id, hours_since_access=hours_since(m.last_accessed),
        importance=importance(m), relevance=relevance(query, m)) for m in memories
], recency_decay=0.995, limit=20)
decision = decide_from_evidence_notes([
    EvidenceNote(document_id=n.document_id, relevant=n.relevant,
        supports_answer=n.supports_answer) for n in model_notes
])