Uteke

Uteke

Long-term memory for any AI agent.

One Rust binary, runs offline, no API key. Self-host it free, or let us host it on Uteke Cloud .

Quick start · Benchmarks · Connect your agent · Docs · Uteke Cloud · Bahasa Indonesia

All 500 LongMemEval-S questions, no LLM call in retrieval. Raw outputs are committed: check them in a minute, or re-run everything for about $10 .

---

Why Uteke

On Monday you tell your agent how you deploy. By Thursday it has no idea. Every session starts from zero, and every agent keeps its own half-memory.

Uteke gives all of them one long-term memory that runs on your own machine:

• Verifiable recall. 98.2% recall@5 on LongMemEval-S, with the stricter numbers shown next to it and the raw outputs in this repo.
• Fast and cheap to query. ~45 ms warm recall, 0 LLM tokens per query, CPU only.
• One binary. No Python, no Docker, no vector database, no API key. The embedding model (~200 MB) downloads once.
• One memory, every agent. Claude Code, Claude Desktop, Cursor, Codex, Hermes and any MCP client share the same store. Rooms record which agent or person wrote what.
• Finds by meaning and by keyword. Local vectors and SQLite full-text search, fused into one ranking (how it works).
• Remembers over time. Point-in-time recall (--at), a relationship graph, a document engine, and decay for stale memories.
• Private by default. No telemetry. Secrets such as API keys are masked before a memory is stored.
• Open source. Apache-2.0, cosign-signed releases.

---

Quick start

# Install (macOS Apple Silicon, Linux; Windows: see below)

curl -sSL codecora.dev/uteke/install | sh

Store a memory

uteke remember "Deploy v2.1 to staging at 3pm"

Search it back: by meaning AND by keyword

uteke recall "when do we deploy?"

Look at what is stored, in a read-only browser page

uteke view

No API keys, no Python, no cloud account. The first run downloads the embedding model (~200 MB, one time).

Give your agent (Claude Code, Cursor, Hermes) the same memory with one line:

// .mcp.json

{ "mcpServers": { "uteke": { "command": "uteke-mcp" } } }

New to Uteke, or you *are* an AI agent? Run uteke onboard. It detects your setup, asks which agent you use, and wires it up. 📖 Onboarding

Rather not run anything locally? [Uteke Cloud](https://uteke.app/?utm_source=github&utm_medium=readme&utm_campaign=prove-the-recall) runs the same engine for you: add one URL to Claude, sign in, done. Free during the open beta.

📦 More install options & Docker

📖 Full install guide · Docker docs

---

Self-host or Cloud

Same engine, two ways to run it.

Choose self-host if your data must stay on your machine, or you want to run Uteke inside your own infrastructure. Choose Cloud if you want one memory across several devices and agents without running a server.

---

Benchmarks

LongMemEval-S (ICLR 2025) hides the facts an agent needs across ~115 chat sessions per question and checks whether retrieval finds them. 500 hand-curated questions, five memory abilities. Uteke runs the full suite with zero LLM calls in the retrieval path: local embeddings, one CPU, deterministic.

¹ Their published numbers on the same benchmark and 500-question split, on their own recall_any@5 basis (head-to-head notes); evaluation setups differ. ² Strict = *every* gold session in the top 5; 65% of questions have more than one. Max possible 99.4%. ³ agentmemory publishes MRR 88.2%, which is a different metric.

Stability across releases and results per question type

Stable across releases (full 500 questions, scored from committed raw, RESULTS.md):

Re-validated 2026-11-08 after the #1690 import fix: strict@5 identical (88.0%), strict@10 +1 question (95.4%), zero misses@50 — four questions shifted partial gold ordering, none lost all evidence.

By question type (recall_any@5, v0.20.0):

Every question's evidence lands in the top 50 (zero misses). What's left is ordering when a question needs several sessions at once: strict recall_all@5 is 88.0% against a 99.4% ceiling, and most misses lose exactly one session.

🎯 Don't trust our benchmark. Run it yourself. Public dataset, committed raw outputs for every validated release, scoring you can recompute in ~20 lines of Python. Verifying the committed results takes a minute and needs no embedder; re-running all 500 questions costs about $10.

👉 [benchmarks/longmemeval/REPRODUCING.md](benchmarks/longmemeval/REPRODUCING.md) · 🧾 Receipts for the headline number (v0.19.0 run)

LoCoMo. LoCoMo end-to-end QA measures answer accuracy, and the result depends on the LLM answerer and the LLM judge used. In an internal comparison, identical answers were scored up to 8 points apart when only the judge model was swapped. For that reason we do not publish a single LoCoMo number in this README. The full table, the per-judge figures and the receipts are in benchmarks/locomo/RESULTS.md. Retrieval recall, which needs no judge, remains the metric we rely on.

uteke bench --counts 100,1000,10000 measures latency and throughput on your own machine. 📖 Benchmark docs

---

Connect your agent

🏠 Rooms: several agents, one shared memory (example)

# Create a shared room

ut