Back to Arena

MemPalace

by Ben Sigman / Milla Jovovich (independent open-source)

System Card

OrganizationBen Sigman / Milla Jovovich (independent open-source)
Released2026-04
Architecturehierarchical-summary / Memory Palace Hierarchical Spatial Index
DetailsMemPalace organizes all conversation data into a four-level hierarchy — Wings (projects/people), Halls (memory types), Rooms (specific ideas), and Drawers (verbatim files) — with 30x-compressed AAAK-format Closet summaries alongside verbatim originals in ChromaDB + SQLite. No summarization or extraction is performed on stored content; only the retrieval layer (170-token structural context injection) uses the hierarchy to boost accuracy 34% over flat search. Entirely offline in raw mode (zero API calls); optional LLM reranking with Haiku/Sonnet achieves up to 100% R@5 on LongMemEval.
Parameters—
Domainepisodic-sessionlifelong-learningrag-retrieval
Open SourceYes
WebsiteVisit
open-sourceofflinememory-palacelocal-llmhierarchicalmit-license

Capability Profile

Benchmark Scores

3 of 14 benchmarks
Data Transparency:2 self-reported1 estimated
Long-Context Retrieval
0/5
RULER
no data
NIAH
no data
LooGLE
no data
LongBench
no data
∞Bench
no data
Multi-Turn Recall
2/2
LoCoMo
88.998pSelf-Reported
MemoryBank
80.378pEstimated
Cross-Session Memory
1/1
LongMemEval
96.699pSelf-Reported
Multi-Hop QA
0/3
BABILong
no data
MultiHop-RAG
no data
HotpotQA
no data
Agent Task Memory
0/1
AgentBench-Mem
no data
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:Self-reported R@5 retrieval recall, raw ChromaDB mode (no LLM). Best config (Hybrid v4 + Haiku rerank) achieves 100% R@5. Held-out 450q test: 98.4%.Self-reported R@10 Hybrid v5 top-10 (no LLM). Best config (Hybrid v5 + Sonnet rerank top-50) achieves 100% R@5/R@10. Baseline session-no-rerank: 60.3%.Arena estimate — MemPalace BENCHMARKS.md mentions MemBench (ACL 2025) but no specific score published