memxt

Install

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

~/.memxt/bin/memxt demo     # 60-second tour on a throwaway palace (the GIF above)
~/.memxt/bin/memxt adopt --write   # wire your agents + mine this repo
~/.memxt/bin/memxt doctor   # if anything's off — every ✗ prints its fix command

macOS / Linux, x86_64 & arm64. One ~7 MB binary + an on-device MiniLM model. No accounts, no keys, zero network at query time.

Related MCP server: umo-memory

Claude Code

/plugin marketplace add Yupcha/memxt
/plugin install memxt

Done. Every session now wakes up with project decisions and recent work (~10 ms), auto-saves before compaction and on stop, and recalls via memory_search → memory_get. /remember and /recall when you want to be explicit.

Other agents — one shared palace

All of them read and write the same MEMXT_DB. Open a new session after wiring. Full guide → docs/harnesses.md.

How it works

   Wings (projects) → Rooms (topics) → Drawers (verbatim memory + embedding)
                              │
          search  =  vectors + FTS5 + facts + recency, fused
          wake-up =  identity + project truth + recent work, ~10 ms

Memories are kept verbatim (no LLM rewriting), searched hybrid (semantic + exact keywords like 0x5C), and consolidated by dream: hot vectors stay f32, cold history compresses to 4-bit, stale facts expire.

Numbers (Apple Silicon, method in BENCHMARK.md): wake-up ~10 ms · warm search sub-ms · ~100 MB peak RAM · continuity bench 6/6 · 72% fewer tokens than re-pasting docs each session (method).

Built for agents (v0.4)

• Procedural memory — repeated successful workflows become procedures; memxt skills --emit turns them into Claude Code skills.

• Grounded memory — memories anchor to file + content hash; recall tags [stale] when the evidence changed.

• Token-budget recall — budget_tokens on search/wake-up returns the best brief that fits.

• Fleet-ready — parallel subagents share one palace: attributed writes, lock-retry, scratch tier with memory_promote.

• Usage-learned relevance — memories the agent actually opens rank higher; ignored ones decay.

• Sleep-time compute — memxt dream --daemon consolidates in the background and flags contradicting facts.

• Client-model sampling — opt-in (MEMXT_SAMPLING=1): fact extraction by your client's own model, still $0.

vs the alternatives

We studied 15 agent-memory products (full teardown). The field splits into cloud memory APIs (Mem0, Supermemory), Python/Node local layers (MemPalace, agentmemory, MemOS), and Claude-first plugins (claude-mem). memxt's bet: the only native-binary, $0-memory-loop option that's deep in every coding harness.

Nobody else has: grounded memories that admit they're [stale], procedural memory that emits agent skills, token-budget recall, or a 60-second demo you can verify on your own repo.

CLI

memxt demo [--keep]                  60-second tour on a throwaway palace
memxt doctor                         Self-diagnosis with exact fix commands
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 [--budget N] | inspect | dream | serve
memxt dream --daemon | --status | --contradictions
memxt skills [--emit [dir]]          Procedural memory → Claude Code skills
memxt anchors [--verify]             Grounded-memory anchor health
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

Zig 0.16 + cmake:

git clone --recursive https://github.com/Yupcha/memxt && cd memxt && zig build --release=fast

ROADMAP.md · MIT · 3D logo

If memxt saves you one re-explain session, star it.