Memexa MCP 持久記憶實測:AI 編碼智能體跨工作階段的表現與問題
I gave my AI coding agents persistent memory via MCP — here's what actually happened
作者實測 Memexa 這個透過 MCP 伺服器為 AI 編碼智能體提供持久記憶的工具,確認專案脈絡可跨工作階段保留。初始化只需一條命令,設定精靈可識別 Claude Code、Cursor、Devin、Codex、Gemini 等 Agent;首次更新因舊遷移未加入 Postgres enum 的 SESSION 值而回傳 500,作者約 30 分鐘修復。
My agents forget everything between sessions — decisions, bug root causes, "why we didn't use MongoDB". All gone when the context
window resets.
So I tried Memexa, a persistent memory layer for coding agents exposed as an MCP server.
## Setup: one command
\bash
cd my-project
npx -y memexa-mcp init
\\
Wizard detects your agents (Claude Code, Cursor, Devin, Codex, Gemini…), asks for an API key once, stores it privately. Config
s contain zero secrets — committable to git, teammates get the same setup by cloning.
It also writes usage rules into AGENTS.md/CLAUDE.md so agents call get_context, add_decision, log_bug without being as
ked.
## It broke on me (instructively)
First update_context → 500, retried 4×. Logs showed invalid input value for enum snapshottype: "SESSION" — a Postgres enum c
reated by an old migration never got the new label because a later migration skipped when the column already existed.
ALTER TYPE … ADD VALUE ×4 → fixed in ~30 min. First snapshot written right after.
## What's actually good
-
add_decisionstores the why + rejected alternatives — the stuff diffs don't capture -
log_bugwith root cause → same bug never diagnosed twice - pgvector
search_memoryfinds entries even with different wording - One key → configs for every agent you use
## Honest limits
- Prompt-driven: agents can ignore it. Works with Claude Code/Cursor/Devin; tiny local models won't bother
- Free tier = 2 projects
- Rough edges expected — it's new (I found one in the first hour)
## Verdict
Context now survives sessions. get_context at startup gives agents real project awareness instead of a blank slate.
\bash
npx -y memexa-mcp # key at app.memexa.tech, free tier
\\
Hands-on tryout, not a paid review.
來源:DEV Community · dev.to