memXT

Local long-term memory for coding agents — Claude Code, Codex, Cursor, Grok CLI.
    Your agent stops forgetting. Your code never leaves your machine.

Install

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

Claude Code

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

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…