Reference

Memory organization and evidence notes¶

Behavior¶

Organization policy	LongMemEval observation	Reading
Recency + importance + BM25 relevance	Evidence recall@10 `0.953`. Complete recall `0.906`	Relevance dominates under the measured `0.15/0.15/0.70` weights
A-MEM link thresholds	Precision `0.436`. Recall `0.944`. F1 `0.597`	Broad linking finds evidence and creates many weak links
Chain-of-Note source choice	Gold-judgment component check	Add a measured classifier before making an answer-quality claim

Actual LongMemEval retrieval differences

Question ID	Type	Gold sessions	Gold ranks	Recall-all@5
`e47becba`	Single-session user	1	`2`	1.0
`6d550036`	Multi-session	4	`4`, `8`, `>10`, `>10`	0.0

The second case retrieves one supporting session near the top. The complete evidence set remains missing. Recall-any records the early hit, so inspect Recall-all for this failure.

The functions link related notes and rank memories for recall. Another function decides 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_kit.knowledge import (
    MemorySignal, plan_note_evolution, rank_salient_memories,
)
from mari_kit.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
])

Authorize the candidate set before ranking or note linking. Min-max normalization is relative to that set, so adding candidates can change existing scores. Treat salience as a ranking signal within one query. Similarity-based links propose related notes. Source agreement and independent corroboration require separate evidence checks.

For cached note metadata or summaries, declare source membership, source revisions, and model configuration through the shared dependency planner. Preserve exact evidence bindings even when an unchanged text representation is reusable.