v0.3.0rc1 — Local-first, zero-LLM agent memory

Memory that stays yours.

Local-first agent memory with zero LLM in the retrieval path. No cloud. No token cost. No rephrasing. Your words, returned verbatim — every qualifier, every detail, intact.

Get started → See how it works
Scroll
01 Architecture

Zero LLM
in the retrieval path.

Every other memory tool calls an LLM when you retrieve. Memex doesn't. BM25 + dense vectors + reciprocal rank fusion. Pure math, zero inference tokens, zero latency tax.

# Default retrieval — zero LLM calls, zero cost from memex import memex client = memex("http://127.0.0.1:19420") result = client.recall("what did we decide about auth?") # → 83.2% R@1, $0.00 per query, 90ms avg latency # → No model API call. No token consumption. No privacy risk.
✓ BM25 + Dense + RRF ✓ 90ms p50 latency ✓ $0 per query ✓ No model API needed
02 Fidelity

Verbatim text.
Never rephrased.

You stored "auth tokens expire after 3600 seconds. The window is non-configurable." Other tools return "authentication has a configurable timeout." Memex returns your exact words — the qualifier survives, every detail intact.

# You stored: "auth tokens expire after 3600s. Non-configurable." # mem0 returns: "authentication has a configurable timeout" # ← the "non-configurable" detail is GONE # Memex returns: "auth tokens expire after 3600s. Non-configurable." # ← your exact words. Every qualifier preserved.
✓ Integer-pointer fidelity ✓ No LLM rephrasing ✓ No hallucination risk ✓ Correction history
03 Privacy

Your data
never leaves.

Memex binds to 127.0.0.1. No cloud account. No third-party processor. No API keys to manage. Your memories live in ~/.memex/ — on your machine, under your control.

# Your data path — fully local, fully yours ~/.memex/ store/ # ChromaDB + SQLite (vector store) graph.db # Knowledge graph (SQLite, no Neo4j) audit.log # Access trail (if MEMEX_AUDIT_LOG=true) backups/ # Auto-compressed daily backups sync/ # Cross-machine sync state # Security layers (all optional, all opt-in): # API token auth, rate limiting, encryption at rest, # audit logging, CORS hardening, request validation
✓ Loopback only (127.0.0.1) ✓ No cloud account ✓ No telemetry ✓ Encryption at rest ✓ Audit logging
04 Economics

$0 per query.
$0 per 1,000 queries.

mem0 costs $2 per 1,000 queries. Zep costs $3. Letta costs $4. Memex costs nothing on the default path — no LLM call means no inference cost. Scale to millions of queries without a billing surprise.

# 1,000 queries — cost comparison mem0: $2.00 (2 LLM calls per query) zep: $3.00 (3 LLM calls, requires Neo4j) letta: $4.00 (every op = LLM tool call) memex: $0.00 (0 LLM calls, pure math) # 10,000 queries/day for a year: mem0: $7,200/yr memex: $0/yr # ← that's a person's salary
✓ No metered pricing ✓ No 13× price cliff ✓ No credit system ✓ Scale to millions for $0
05 Platform

4 SDKs. 25+ CLI tools.
One memory layer.

Memex works with every agent: Claude Code, Cursor, Codex, Windsurf, Continue.dev, Zed. SDKs in TypeScript, Python, Go, and Rust. CLI with doctor, backup, graph, TUI, benchmark, audit, and more.

# CLI commands $ memex init / watch / recall / query / add / health $ memex doctor # diagnose prereqs + runtime $ memex backup export # backup to compressed JSON $ memex ui / tui # web UI + terminal UI $ memex graph auth # knowledge graph query $ memex stats # cost comparison vs competitors $ memex benchmark # run LongMemEval-S $ memex audit --stats # audit log analytics $ memex sync --push # sync across machines $ memex encrypt --status # encryption at rest # SDKs import memex # Python import { memex } from '@alicelabs/memex-sdk' # TypeScript import "github.com/eddyflores100-lang/memex/sdk/go/memex" # Go use memex::MemexClient; # Rust
✓ TypeScript SDK ✓ Python SDK ✓ Go SDK ✓ Rust SDK ✓ MCP integration ✓ MEMEX-MEMORY/v1 standard
0
Retrieval R@1
0
Cost per query
0
Avg latency
0
LLM calls on default path

How we compare

No cloud. No LLM tax. No vendor lock-in.

mem0 Zep Letta Memex
Local-first — — — ✓
Zero LLM in retrieval — — — ✓
Verbatim passages — — — ✓
Cost / 1K queries $2 $3 $4 $0
R@1 (independent) 49% 63.8% — 83.2%
Graph memory — Neo4j — SQLite
SDKs Python Python Python 4 languages
External deps Cloud Neo4j + LLM LLM None
Compliance docs — — — SOC2, HIPAA, GDPR

"A summary can preserve 'we tried the migration' while dropping why it failed. Memex keeps those details in the stored note — not a generated fact."

Design Principle — Verbatim Fidelity

Get started

Install in
60 seconds.

Python 3.10+, macOS or Ubuntu, Ollama running locally. That's it.

# 1. Install pip install -e . # 2. Start the service memex init # 3. Ingest your notes memex watch ~/notes # 4. Recall memex recall "what did we decide about auth?" # 5. Explore memex ui # web UI on :19421 memex tui # terminal UI memex graph auth # knowledge graph memex stats # cost comparison vs competitors
Read the docs → View on GitHub

Your agent forgets everything.

Stop re-explaining context. Stop losing decisions between sessions. Give your agent memory that stays.

Get Memex → Compare alternatives