🪶 Feather DB — Living Context Engine

Embedded vector DB · HNSW search · typed context graph · adaptive decay · MCP server
getfeather.store · Hawky.ai · PyPI · GitHub · pip install feather-db

📊 Latest benchmark — LongMemEval (Apr 2026): Feather + GPT-4o = 0.693 · Feather + Gemini-Flash = 0.657 — beats the LongMemEval paper's full-context GPT-4o ceiling (0.640). Reproducible on a $2.40 budget. Full report →

Demo graph: 10 nodes representing product intelligence for an AI developer tools team — feature performance, competitor moves, community signals, strategy briefs, and user research. 8 typed causal edges connect them (contradicts, supports, derived_from, references).

Feather DB v0.8.0 — Reproducible Benchmark Results

LongMemEval (Xu et al., 2024 / ICLR 2025)

500-question end-to-end memory QA benchmark. Each question carries up to ~115K tokens of chat history; system must ingest, retrieve, and answer correctly across 5 memory ability axes.

System Variant Answerer Overall Cost / run
Feather DB v0.8.0 + decay S gpt-4o 0.693 ~$8
Feather DB v0.8.0 + decay S gemini-2.5-flash 0.657 ~$2.40
Full-context GPT-4o (paper "ceiling") S gpt-4o + CoN 0.640 n/a
Zep (graphiti) S gpt-4o-mini 0.638 (vendor)
Full-context GPT-4o-mini S gpt-4o-mini 0.554 n/a
Naive vector RAG (paper) S/M gpt-4o ~0.31 n/a

Feather + GPT-4o (0.693) beats the LongMemEval paper's full-context GPT-4o ceiling (0.640). Our 10-snippet retrieval carries more useful signal to the answerer than dumping the whole 115K-token haystack into a frontier model — at ~40× lower input-token cost per query.

Per-axis (Feather + GPT-4o on S)

Axis Score
single-session-user 1.000 (perfect)
single-session-assistant 0.964
single-session-preference 0.767
knowledge-update 0.714
multi-session-reasoning 0.606
temporal-reasoning 0.477

ANN performance — SIFT1M (real data, 500K × 128-dim)

ef p50 latency p99 latency Recall@10
10 0.07 ms 0.13 ms 0.774
50 (default) 0.19 ms 0.23 ms 0.972
100 0.32 ms 0.39 ms 0.991
200 0.56 ms 0.69 ms 0.998

Reproduce

pip install feather-db
git clone https://github.com/feather-store/feather && cd feather

python -m bench run longmemeval --dataset s --limit 0 \
    --embedder openai \
    --answerer-provider gemini --answerer-model gemini-2.5-flash \
    --judge llm --judge-provider gemini --judge-model gemini-2.0-flash \
    --decay-half-life 14 --decay-time-weight 0.4 --k 10

Audit trail

Per-run JSON results — every number above is one of these files:

Connect Feather DB to any LLM in 5 lines:

pip install feather-db

from feather_db.integrations import ClaudeConnector
conn   = ClaudeConnector(db_path="my.feather", dim=3072, embedder=embed_fn)
result = conn.run_loop(client,
    messages=[{"role": "user", "content": "Why is onboarding drop-off so high?"}],
    model="claude-opus-4-6")

Works with Claude · OpenAI · Gemini · Groq · Mistral · Ollama · MCP (Claude Desktop, Cursor)

getfeather.store · Hawky.ai · PyPI · GitHub · Integrations Guide