memoket/memoket-kite
An open-source memory layer that answers questions with structured facts and readable plans instead of vector similarity
memoket-kite stores conversations as structured, dated, topic-tagged facts linked back to the original message, rather than turning everything into embedding vectors. A question is compiled into a human-readable query plan that is executed over this fact store, and answers come with the exact source line and timestamp, or an honest empty result if nothing was ever said. Evaluated with gpt-4.1-mini as the shared reader and judge, it tops LoCoMo at 93.51% and LongMemEval-S at 85.60% while reading only about 1.6K tokens of context per question.
What it does
- Converts messages into typed, dated facts tagged by topic, keeping the original line as evidence rather than discarding it into a vector
- Compiles each question into an inspectable, cacheable JSON plan instead of an embedding search, so you can see exactly which conditions selected which facts
- Runs on a single topic-indexed file with no embeddings, vector database, or reranking layer required
- Can resolve conflicting statements made at different times by sorting on event time, so a question about 'now' returns the more recent fact
- Ranked first on both LoCoMo (93.51%) and LongMemEval-S (85.60%) under one shared evaluation protocol using gpt-4.1-mini as reader and judge, with an average reader context of about 1.6K tokens per question
Why it matters
AI agents that need to remember long conversations can pull the fact that actually answers a question, not just the sentence that sounds similar, and can show the receipts for why that answer was chosen. Running without a vector search stack, on a lightweight portable file, also matters for teams weighing reproducibility and infrastructure cost.
Terms in this repo
- embedding · a numeric vector representation of text used to compute semantic similarity
- vector database · a data store optimized for finding the nearest embedding vectors to a query
- query plan · a human-readable, structured description of how a question is turned into filter and sort steps over the fact store
- LoCoMo, LongMemEval-S · public benchmark datasets that test memory over long conversations
- reader context · the average amount of evidence text the answering model actually gets to see per question
Repository description (English)
Memory layer for AI agents with token-efficient, explainable retrieval beyond vector similarity.
Open on GitHubTrending repos
- vorssaint/vorssaint-utilsOne free menu bar app replaces a dozen paid Mac utilities
- Alishahryar1/free-claude-codeA local proxy that lets coding AI agents run on 49 free or cheap model providers instead of one paid service
- freestylefly/awesome-gpt-image-2A library of 532 reverse-engineered prompts that turn GPT-Image2 into a predictable image-making tool
- block/buzzAn open-source workspace where humans and AI agents chat, code, and review in the same rooms
- NousResearch/hermes-agentNous Research's Hermes is an AI agent that gets smarter the more you use it
- virgiliojr94/book-to-skillA tool that turns technical book PDFs into on-demand reference skills for AI coding agents
- VoltAgent/awesome-agent-skillsA single hub collecting over 1000 'how-to' manuals that make AI coding assistants act like experts
- anthropics/claude-plugins-communityA shared shelf where anyone's Claude add-ons get listed for install