KITE：为智能体提供可追溯记忆

An open-source, vector-free long-term memory engine for AI agents, achieving SOTA on LoCoMo and LongMemEval with significantly less context.

Follow the thread, not the nearest match. Open-source and vector-free, KITE achieves 🥇 93.51% on LoCoMo and 🥇 85.60% on LongMemEval-S, with gpt-4.1-mini used across all benchmark stages and an average reader context of only ~1.6K tokens per question. 📢 News [2026-10-26]: 🎬 Released an explainer video of how KITE works, from a message to a cited answer. [2026-08-13]: 🎉 First public release: KITE v0.1.0 is open source under Apache-2.0 and on PyPI as memoket-kite. 💡 Why KITE by Memoket Most AI memory turns every sentence into a cloud of numbers and hands back whatever lands closest to your question. It always returns something, even when the answer was never there, and it cannot show you why. KITE by Memoket reads what a question means, not how it sounds. "Who's the decision-maker on the Henderson account now?" Nothing in memory says decision-maker: March. Marcus: "I'll sign off on Henderson." June. The new VP of Ops: "Dana owns Henderson from here." Both were filed under the same topic when they were written. Then now sorts them by date. Similarity has no notion of now: nearest-match memory rates March just as relevant and answers "Marcus", with nothing to tell you it is wrong. KITE answers "Dana", with the June line attached. Nearest-match memory 🪁 KITE by Memoket Finds whatever sounds like the question the facts that answer it When it was never said returns something anyway comes back empty, and says so Why this answer? a similarity score a readable plan, with who said it and when Runtime stack embeddings · vector DB · rerankers one portable, topic-indexed file Ask again tomorrow depends on the index that day the same plan walks the same steps ⚙️ How It Works The whole idea in one line: conversations settle into facts under topics, and a question finds its topic and pulls that topic's memories, receipts included. KITE-how-it-works.mp4 1 · Message → Fact A user says, "I moved to Tokyo on July 26." KITE by Memoket extracts what happened and files it under its topics, keeping the original message as evidence: <fact t="2026-07-26" kind="event" who="user" place="tokyo" event="move" src="chat1L1"> The user moved to Tokyo on 26 July 2026. </fact> <line id="chat1L1" who="user">I moved to Tokyo on July 26.</line> 2 · Question → Plan When the agent asks "Where does the user live now?", KITE by Memoket compiles the question into a plan you can read, validate, and cache: { "select": "facts", "where": {"who": ["user"], "event": ["move"]}, "pipe": [ {"op": "sort", "key": "t", "desc": true}, {"op": "head", "n": 1} ] } 3 · Plan → Execution The executor runs the plan over the topic index. Keywords and associations come first, then conditions filter down to the one moment that answers the question: SEARCH facts FILTER who = "user" AND event = "move" SORT BY event_time DESC TAKE 1 The same validated plan always follows the same steps over the same memory, with no embeddings and no vector database anywhere. 4 · Evidence → Answer Answer: Tokyo Source: chat1L1, "I moved to Tokyo on July 26." Key Features 🚫 No vector stack No embeddings, no vector database, no layers to embed, sort, and re-rank every result. 🗂️ Structured facts Conversations crystallize into typed, dated, topic-indexed memory you can open and read. 🔍 Inspectable queries Every question becomes a symbolic plan, revealing exactly how each answer was matched. 🕐 Time-aware retrieval People, events, dates, and updates handled directly, so "now" means now. 🧾 Shows its work Evidence links straight back to the moment it came from: who said it, when, and the actual words. 🤷 Knows what it doesn't know When something was never said, the search comes back genuinely empty, and KITE says so. 🚀 Quick Start 1. Install and configure python -m pip install memoket-kite export OPENAI_API_KEY="sk-..." mkdir -p artifacts curl -fsSL \ https://raw.githubusercontent.com/memoket/memoket-kite/main/examples/data/demo_codebook.xml \ -o artifacts/quickstart.xml Using another OpenAI-compatible provider? Set OPENAI_BASE_URL as well. 2. Remember Turn a conversation into structured, source-backed facts: from memoket_kite import Memory memory = Memory.load("artifacts/quickstart.xml") memory.remember( [{"role": "user", "content": "I moved to Tokyo."}], session_id="session_1", ) 3. Recall Compile a question into a plan and retrieve matching facts, each carrying its sources: from memoket_kite import Memory memory = Memory.load("artifacts/quickstart.xml") for fact in memory.recall("Where did the user move?"): print(fact.content) print(fact.sources) 4. Answer Generate an answer from the retrieved evidence and keep the receipts: from memoket_kite import Memory memory = Memory.load("artifacts/quickstart.xml") result = memory.answer_with_evidence("Where did the user move?") print(result.text) # "The user moved to Tokyo." for fact in result.evidence: print(fact.content, fact.sources) # the moments the answer stands on print(memory.answer("Where did the user move?")) # just the text Full reference in docs/api.md; more runnable examples in examples/. 📊 Benchmark Results KITE by Memoket posts the top overall score on both public long-conversation memory benchmarks. It is the only system up there that uses no vectors, and it reads less context than any capable rival. Benchmark Overall Accuracy Avg. Reader Context LoCoMo 🥇 93.51% 1.51k tokens LongMemEval-S 🥇 85.60% 1.65k tokens All systems are evaluated under one shared protocol, with gpt-4.1-mini as the common reader and judge; the exact rubrics are in the LoCoMo protocol and LongMemEval protocol. Reader context is the mean per-question evidence budget: what the reader actually gets to see, measured by the same ledger on every question. Each figure comes from a reference run under the default configuration, with gpt-4.1-mini reading and judging and no ablation flags set. The reproduction contract pins the dataset revisions, protocols, model IDs, expected row counts, and reported scores. The provided scripts rebuild the evaluation locally fro