memory-kg 0.9.0

MemoryKG — A Hybrid Knowledge Graph for Conversational Memory and Document Corpora

Author: Eric G. Suchanek, PhD — Flux-Frontiers, Liberty TWP, OH

TL;DR

MemoryKG achieves 96.3% tier-1 retrieval recall on the ConvoMem benchmark at k=20 (500 tier-1 items; 88.7% averaged across all 1,897 items spanning tiers 1–4), exceeding MemPal's published 92.9% tier-1 baseline by +3.4 pp. No LLM, no API key, no cloud inference at any stage. Full write-up: benchmarks/convomem/convomem_article.pdf.

Recall is measured by substring containment in the top-10 retrieved nodes: an evidence message counts as found if its text appears verbatim in (or contains) any retrieved node — lenient toward retrieval, but it cannot be fooled by paraphrase.

On the LongMemEval-S benchmark, MemoryKG scores 98.2% Recall@5, 99.2% Recall@10, 0.954 NDCG@10 with zero inference (re-verified 2026-08-26). Against MemoryPalace hybrid v2 — the best LLM-free configuration either system has — it wins at depth (NDCG@10 0.954 vs 0.934, R@10 99.2% vs 99.0%) and trails by 0.2 pp at R@5 (98.2% vs 98.4%). LLM-augmented systems still rank higher at R@5 (MemoryPalace v4 + Haiku at 100%, v3 + Haiku rerank at 99.4%, Supermemory ASMR at ~99%); MemoryKG closes most of that gap without paying the inference cost. Full write-up: benchmarks/longmemeval/longmemeval_article.pdf.

With the sibling boost enabled on LongMemEval, recall_all@10 reaches 98.6% — meaning MemoryKG retrieves every required session for 493 of 500 questions without any LLM. No published system reports this metric; we track it because multi-session coverage is the real test of memory completeness.

The field has been over-engineering retrieval. A graph-augmented index with correct search-space scoping matches the best LLM-free result in the field at a fraction of the complexity.

Why It Works

Most "memory" systems flatten a session into a single embedding and lean on an LLM at query time to rerank what they retrieve. MemoryKG does the opposite: it preserves session structure as a typed graph, then uses that structure as the ranking signal.

1. Finer granularity. Sessions are chunked by heading, not embedded as 2,000-word blobs. A 150-word chunk about "Dr. Chen's appointment" is dramatically more discriminative than the session it lives in.
2. Structural expansion. A HAS_TOPIC or MENTIONS_ENTITY edge from a weakly-matching chunk surfaces strongly-linked neighbors that pure cosine similarity never finds.
3. Score-first ranking. Graph proximity breaks ties within a vector-quality band — never across one. Good seeds get amplified; bad seeds don't get rescued.
4. Kind-aware ranking. Chunk matches outrank entity stubs outrank synthetic topic summaries. Flat vector stores treat every document equally.
5. Search-space scoping. When the benchmark defines a per-question candidate pool, MemoryKG honours it (haystack_files=...). This was the +11 pp fix that narrowed the gap to the inference-based leaderboard.

No LLM. No API key. No cloud round-trip. Runs on Apple Silicon (MPS), CUDA, or CPU.

What MemoryKG Is

A deterministic, explainable knowledge graph built from conversational logs and document corpora (Markdown, plain text). MemoryKG semantically chunks text, extracts topics/entities/keywords, links them through typed edges, stores everything in SQLite, and adds a sqlite-vec vector index as an acceleration layer — not the source of truth.

Structure is treated as ground truth. Semantic search is a tool, not the system. The result is a searchable, auditable representation that supports precise navigation, source-grounded passage extraction, and downstream LLM reasoning — a practical foundation for Knowledge-Graph RAG (KGRAG).

MemoryKG shares its architecture with PyCodeKG (Python codebases) and DocKG (general document corpora).

Features

• Semantic chunking — Multiple strategies: heading (one chunk per ## Section), fixed, sentence_group, semantic (embedding-boundary detection)
• Deterministic knowledge graph — SQLite-backed canonical store with typed nodes and provenance-tracked edges
• Relation extraction — Topics, named entities, keywords; co-occurrence and similarity edges built automatically
• Hybrid query model — Semantic seeding (sqlite-vec) + structural expansion (graph traversal) + score-first ranking
• Haystack-scoped search — Restrict vector seeding to a per-question candidate pool for benchmark-grade precision
• Passage packing — Source-grounded text passages with headings, ready to paste into an LLM prompt
• Coverage analysis & temporal snapshots — Per-document metrics, hot chunks, orphan detection, version-over-version diffs
• Parallel ingestion — --workers N parallel Phase 1 parsing for large corpora
• MCP server — Four tools for AI agent integration (graph_stats, query_docs, pack_docs, get_node)
• Streamlit web app — Interactive graph browser, hybrid query UI, and passage pack explorer

Quick Start

# Index a corpus (SQLite + sqlite-vec in one step; wipe is the default)
memorykg build --repo docs/

# Natural-language query — returns ranked chunks
memorykg query "authentication flow"

# Source-grounded passage pack — paste straight into an LLM prompt
memorykg pack "configuration reference" --fmt md --out context.md

Installation

pip install 'memory-kg @ git+https://github.com/Flux-Frontiers/memory_kg.git'

See docs/installation.md for editable installs, dev setup, and offline model caching.

Usage

Build the knowledge graph

# Full pipeline: parse → SQLite graph → sqlite-vec index (wipe is default)
memorykg build --repo docs/

# Granular steps for large corpora
memorykg build-graph --repo docs/   # SQLite only
memorykg build-index                 # vector index from existing SQLite

# Incremental update — keep existing data
memorykg build --repo docs/ --update

# Parallelise Phase 1 parsing
memorykg build --repo docs/ --workers 8

# Exclude directories
memorykg build --repo docs/ --exclude-dir archive --exclude-dir vendor

Query and pack passages

# Hybrid query — semantic seed + graph expansion
memorykg query "deployment configuration"

# Tune top-K and expansion hops
memorykg query "API authentication" --k 12 --hop 2

# Pack as Markdown for LLM context injection
memorykg pack "error handling strategies" --fmt md --out context.md

Analyze, snapshot, visualize

memorykg analyze --repo docs/                      # corpus health report
memorykg snapshot save 0.4.1 && memorykg snapshot diff 0.4.0 0.4.1
memorykg viz                                       # Streamlit graph browser
memorykg mcp --repo docs/                          # MCP server for AI agents

See docs/cli-reference.md for every flag.

Reproducing the Benchmarks

LongMemEval-S — 98.2% R@5, 99.2% R@10

Full write-up: benchmarks/longmemeval/longmemeval_article.pdf

# 1. Install
poetry install

# 2. Download LongMemEval-S
mkdir -p /tmp/longmemeval-data
curl -fsSL -o /tmp/longmemeval-data/longmemeval_s_cleaned.json \
  https://huggingface.co/datasets/xiaowu0162/longmemeval-cleaned/resolve/main/longmemeval_s_cleaned.json

# 3. Build the corpus + KG (BGE-small-en-v1.5, heading chunks).
#    Parse -> SQLite, embed -> JSONL cache, index from cache; --keep-cache
#    resumes an interrupted build without re-embedding.
poetry run python3 benchmarks/longmemeval/longmemeval_memkg.py prepare \
  /tmp/longmemeval-data/longmemeval_s_cleaned.json \
  --wipe --chunk-strategy heading

# 4. Run evaluation (haystack filter and k=50 are now defaults)
poetry run python3 benchmarks/longmemeval/longmemeval_memkg.py run \
  /tmp/longmemeval-data/longmemeval_s_cleaned.json \
  --out benchmarks/longmemeval/results_20260826_bge_haystack.jsonl

# Expected: R@1=90.4%  R@5=98.2%  R@10=99.2%
#           recall_all@10=98.8%  NDCG@10=0.954  misses@10=4/500

ConvoMem — 96.3% Tier-1 Recall Across 1,897 Items

Full write-up: benchmarks/convomem/convomem_article.pdf

# Run all four evidence tiers (top-10, hop=1, BGE-small-en-v1.5)
poetry run python3 benchmarks/convomem/convomem_bench.py --limit 1000 --tier 1
poetry run python3 benchmarks/convomem/convomem_bench.py --limit 1000 --tier 2
poetry run python3 benchmarks/convomem/convomem_bench.py --limit 1000 --tier 3
poetry run python3 benchmarks/convomem/convomem_bench.py --limit 1000 --tier 4

# Expected: 96.3% tier-1 recall@20 (500 items); 88.7% averaged across all 1,897 items in tiers 1-4 (~20 min)

Hardware tested: Apple M5 Max MacBook Pro, 64 GB RAM. Also runs on CUDA and pure CPU (MEMORYKG_DEVICE=cpu).

Documentation

Citation

If you use MemoryKG in your research or project, please cite it:

APA

Suchanek, E. G. (2026). MemoryKG: Hybrid Semantic-Graph Knowledge Base for Conversational Memory (Version 0.9.0) [Software]. Flux-Frontiers. https://github.com/Flux-Frontiers/memory_kg

BibTeX

@software{suchanek_memory_kg,
  author    = {Suchanek, Eric G.},
  title     = {{MemoryKG}: Hybrid Semantic-Graph Knowledge Base for Conversational Memory},
  version   = {0.9.0},
  year      = {2026},
  publisher = {Flux-Frontiers},
  url       = {https://github.com/Flux-Frontiers/memory_kg},
  doi       = {10.5281/zenodo.21282909},
}

License

Elastic License 2.0 — free for non-commercial and internal use; commercial hosting or redistribution requires a license from Flux-Frontiers.