-
Notifications
You must be signed in to change notification settings - Fork 2
research(memory): TiMem temporal-hierarchical memory tree — complexity-aware recall, no over-summarization (arXiv:2601.02845) #2262
Copy link
Copy link
Open
Labels
P3Research — medium-high complexityResearch — medium-high complexitymemoryzeph-memory crate (SQLite)zeph-memory crate (SQLite)researchResearch-driven improvementResearch-driven improvement
Description
Summary
TiMem (arXiv:2601.02845, Jan 2026) — Temporal Memory Tree (TMT) that organizes conversational history into hierarchical levels: raw observations at leaves, progressively abstracted persona representations at higher levels. Consolidation is driven by semantic similarity (no fine-tuning required). Recall uses a complexity-aware selector that balances precision vs. latency based on query difficulty.
Applicability to Zeph
Zeph's zeph-memory handles summarization and temporal decay as a flat pipeline. TiMem's tree structure addresses two known gaps:
- Over-summarization of recent events: current compaction can collapse recent turns that should stay granular — tree leaves preserve them
- Under-retention of recurrent persona signals: persona facts that appear across multiple sessions get diluted; TMT's upper levels stabilize them
- Complexity-aware recall: aligns with Zeph's multi-model design principle — cheap model (gpt-4o-mini) for simple leaf lookups, quality model for deep consolidation tree traversal
Implementation Sketch
- Schema: add a
memory_treeSQLite table withlevel,parent_id,content,consolidated_atfields - Consolidation: run as a background sweep (similar to existing tier promotion loop in
zeph_memory::tiers) — group semantically similar recent turns and consolidate to parent nodes - Recall: route queries by complexity (keyword → leaf lookup; semantic → tree traversal up to level N)
- Complexity: MEDIUM — tree schema + consolidation scheduler; no model changes needed
Related Issues
- Complements research: MAGMA multi-graph memory (semantic + temporal + causal + entity) for richer retrieval (arXiv:2601.03236) #2231 (MAGMA multi-graph memory) — MAGMA adds relation-type graphs; TiMem adds temporal hierarchy on the message level
- Complements research: Focus autonomous context compression (agent-driven, 22.7% token reduction) (arXiv:2601.07190) #2218 (Focus autonomous context compression) — TiMem prevents the need for aggressive compression by retaining selective granular history
Source: https://arxiv.org/abs/2601.02845
Reactions are currently unavailable
Metadata
Metadata
Assignees
Labels
P3Research — medium-high complexityResearch — medium-high complexitymemoryzeph-memory crate (SQLite)zeph-memory crate (SQLite)researchResearch-driven improvementResearch-driven improvement