memxt

The 3 a.m. feeling

You explained the architecture yesterday. The cart cap is 37. You banned Redis for a reason. SQLite won the debate.

Today your agent knows none of it. So you paste the README again. Re-explain the constraints again. Watch it reintroduce Redis… again.

Cloud "memory layers" fix this by uploading your code and your chat to someone else's server. Hard pass.

memxt gives your agent a memory palace on your own disk — one it searches and writes as it works, wired straight into the coding loop with MCP + hooks. Verbatim. Local. Yours.

Related MCP server: OpenExp

⚡ Install

curl -fsSL https://raw.githubusercontent.com/Yupcha/memxt/main/install.sh | bash

→ ~/.memxt/bin/memxt + an on-device MiniLM model. macOS / Linux, x86_64 & arm64. No keys, no accounts.

~/.memxt/bin/memxt adopt --write   # wire up your agents + mine the current repo
~/.memxt/bin/memxt inspect         # palace health
~/.memxt/bin/memxt serve           # browse it → http://127.0.0.1:8765

🪄 Claude Code in 30 seconds

/plugin marketplace add Yupcha/memxt
/plugin install memxt

That's the whole setup. From now on, every session just… remembers.

~/.memxt/bin/memxt mine . my-project   # optional: seed the palace from your repo

Next session, nobody re-explains the 37-item cart cap, the SQLite choice, or the Redis ban. It's just known.

🧩 Every agent, one palace

One MEMXT_DB. Every subscription you already pay for reads and writes the same memory.

~/.memxt/bin/memxt instructions --harness grok   # also: codex · cursor · zed · claude

Standing rules live in AGENTS.md so agents know when to reach for memory — not just how. Full multi-harness guide (doctor, restart, troubleshooting) → docs/harnesses.md.

MCP loads at session start. After adding memxt, open a new session.

✅ Proof, not vibes

Coding Continuity Bench — 6/6. Paraphrased questions (cart 37 / 0x5C, SQLite vs Postgres, the Redis ban, HttpOnly cookies) + wake-up + profile. Every one recalled.

./scripts/bench-continuity.sh

Fast where it counts — Apple Silicon · Metal:

Method → BENCHMARK.md · details → docs/BENCHMARKS_CONTINUITY.md.

📉 Token savings

Not a synthetic leaderboard. On this repo: wire up the agent, mine the project, ask 5 continuity questions.

paste the docs every restart    ████████████████████████████  9,623 tokens
memxt wake + targeted recall     ████████                      2,703 tokens

* tiktoken cl100k_base. Honest caveat: a perfect 176-token hand-written recap can look cheaper — but real amnesiac sessions re-dump docs or just fail. Full method → docs/TOKEN_SAVINGS.md.

🏛 How it works

                a palace on your disk
                        │
   Wings (projects) → Rooms (topics) → Drawers (verbatim memory + embedding)
                        │
        search  =  vectors + FTS5 + facts + recency, fused
                        │
        wake-up =  who you are + project truth + recent work,  in ~10 ms

memxt search "cart limit" --mode hybrid
memxt search "SQLite" --mode facts --wing my-project
memxt dream --budget 50000        # compress old memories, keep them searchable
memxt serve --port 8765           # inspect / search / dream in the browser

How it stacks up

We don't chase cloud LoCoMo / LongMemEval leaderboards. We win the thing you actually feel: your agent stops forgetting, and your code stays on your machine.

The closest peer is claude-mem — genuinely great at automatic Claude capture. If you've hit its cloud-compression costs, or you also live in Grok and Codex, memxt is the same continuity with zero memory tax and one palace across all of it.

Full teardown (feature kill-sheet, stack matrix, 8 sharp tools vs 53) → research/COMPETITOR_KILL_SHEET.md.

cloud memory API   ──►  your code + chat leave the building
memxt              ──►  a SQLite palace on disk · 0 API keys · 0 query network

CLI

memxt adopt [--write] [--no-mine]    Wire up agents + optionally mine the repo
memxt mine <path> [wing]             Incremental codebase ingest
memxt search <q> --mode hybrid | memories | documents | facts | episodes
memxt wake-up | inspect | dream | serve
memxt forget | export | import | mcp | hook | instructions

MEMXT_DB=~/.memxt/palace.db
MEMXT_MODEL=~/.memxt/lib/minilm.gguf
MEMXT_WING=my-project    # optional; defaults to the git-root name

Build from source

Most people never need this. Requires Zig 0.16 + cmake:

git clone --recursive https://github.com/Yupcha/memxt && cd memxt
zig build --release=fast   # builds llama.cpp (Metal on macOS) + fetches the MiniLM GGUF

Details live in install.sh and the comments in build.zig.

Roadmap · License · Star

• Roadmap → ROADMAP.md

• Launch notes → docs/launch/

• MIT → LICENSE

↑ drag it in 3D

If memxt saves you one re-explain session, star it and tell another Claude Code user.

That's how local tools win. ⚡