vshulcz (@vshulcz_dev) profile picture
Joined April 2017
Photo by vshulcz (@vshulcz_dev)Photo by vshulcz (@vshulcz_dev)
Glad... now you can have one common shared memory across 34+ AI coding agents and even remember what they did months ago. 🤯 Deja-vu (as the name says) is an open-source memory layer for coding agents that indexes the session history already stored on your machine and lets different agents recall what was done in previous sessions. So a fix you discovered in Claude Code can be recalled by Codex, Cursor, Antigravity, and other supported agents. Basically... your AI coding agents finally have a shared memory :) Follow me for more amazing AI, Coding & Web Dev insights 💎
11
4
19
1,207
An interesting small project called deja-vu takes a very simple approach to coding-agent memory. No LLM, no embeddings, no vector database. It indexes the session history already stored by tools like Claude Code, Codex, Cursor, and Copilot, then makes previous work searchable across agents. github.com/vshulcz/deja-vu
1
2
3
1,349
vshulcz retweeted
A new tool, deja-vu, indexes every transcript your AI coding agent writes, Claude Code, Codex, opencode, and says 'nothing leaves your machine.' A tool with that much access, I had to read the source. Good news: the claim mostly holds. Here's exactly what stays local, what can leave, and 2 caveats.
3
2
15
502
xAI open-sourced Grok Build this week. Its sessions live in ~/.grok/sessions as undocumented jsonl. deja-vu now indexes them — searchable together with your Claude Code / Codex / opencode history, with MCP recall inside every agent. All local. github.com/vshulcz/deja-vu
3
3
73
vshulcz retweeted
Agent对话日志写了几百MB,查个bug还得重放会话? deja-vu 直接索引 Claude Code、Codex、opencode 的本地记录,毫秒级搜索,零依赖二进制。 值得留意的是它不只搜——还能通过 MCP recall 让 agent 自动回忆“之前怎么修的”,甚至 SessionStart 自动注入相关上下文。 敏感信息在索引时就脱敏,同步、分享都安全。 增量索引,只读新文件,跑一次就基本常驻。 说白了,给 coding agent 加了个不联网的长期记忆层。 仓库在这 👇 github.com/vshulcz/deja-vu
1
1
58
vshulcz retweeted
Repo bagus nih buat workflow agent namanya dejavu, simpel dia bikin semua chat dan debug lama dari Claude Code, Codex, sampai OpenCode jadi bisa dicari lagi. Jadi pas error yang sama muncul minggu depan, agent nggak perlu ngulik dari nol atau nanya konteks lama lagi. Tinggal cari kata kuncinya, lalu keluar riwayat masalah, penyebab, dan fix yang pernah dipakai. github.com/vshulcz/deja-vu
1
3
12
634
如果一个编程智能体每次新会话都要重新认识项目,问题未必出在模型能力,而可能出在记忆没有成为独立、可迁移的工作层。 GitHub 上新发布的这个开源项目,尝试用 SSH 同步编程 AI 智能体的记忆数据,让智能体跨会话保留上下文。它不依赖特定云服务,用户可以自托管;代码已经开源,也为开发者留下了集成和定制空间。 这套思路的价值,不是单纯“记住更多”,而是把记忆从某个智能体或某次会话里拆出来。只要同步层能够进入现有工作流,用户就有机会自己掌握记忆数据,并在会话变化时继续使用,而不必完全依赖特定云端产品提供的连续性。 但“能通过 SSH 同步”只回答了数据怎么移动,没有回答什么值得被记住。编程智能体的有效记忆可能涉及项目约定、已做决策、任务状态和历史错误;如果不加筛选地长期累积,旧信息也可能干扰新任务。同步能力与记忆质量是两个不同问题,前者不能自动保证后者。 同样需要区分自托管与低成本。自托管意味着控制权和维护责任都回到用户手里。输入没有说明项目的记忆结构、检索方式、冲突处理、权限设计、数据保护、清理机制或规模表现,因此现在还不能判断它是否适合敏感代码库、多人协作或长期生产环境。 我会用四个问题评估这类方案: 一,记忆是否可读、可查、可删除,而不是只会不断写入; 二,跨会话恢复的究竟是有效决策,还是未经整理的聊天残留; 三,SSH 同步如何进入现有开发流程,会增加多少配置与维护工作; 四,切换智能体或设备后,数据能否继续被实际使用,而不只是完成文件传输。 对个人开发者,较稳妥的采用方式是先在非敏感项目中测试一个边界明确的任务,观察新会话能否正确继承必要上下文。对团队,则应先确定哪些记忆可以共享、谁负责更新,以及错误信息如何纠正,再讨论全面接入。 这个项目值得关注,因为它把编程智能体竞争中的一个隐性问题摆到了台前:模型会不会写代码是一层,长期协作时能不能保留正确上下文是另一层。SSH 同步提供了一个开放、自托管的实现方向;它最终能否降低重复说明成本,要看记忆治理是否与同步机制一起成立。 github.com/vshulcz/deja-vu
2
1
7
917
Este proyecto me encantó: Deja-vu es una herramienta para mantener memoria persistente para tus agentes de AI PERO, el diferenciador entre tantas opciones que he visto: sincronización entre dispositivos github.com/vshulcz/deja-vu/
5
25
149
7,144
Claude Code、Codex、opencode のローカルセッションログを横断検索・MCP recall・SessionStart 自動コンテキスト化するライブラリ。 vshulcz/deja-vu: Memory layer for coding agents: search, MCP recall, auto-context, secret redaction, stats, share and sync over the session logs Claude Code, Codex and opencode already write. One zero-dep binary. github.com/vshulcz/deja-vu/
3
18
2,057
your coding agents already solved that bug. they just can't remember it. deja v0.9 now reads 6 harnesses into one local memory: claude code, codex, opencode, cursor, gemini cli, aider. one zero-dep binary. 7ms search over months of transcripts. no cloud. github.com/vshulcz/deja-vu
1
1
130
the part i actually built it for: `deja sync ssh mini` my laptop's agent knows what the headless mac mini tried overnight - what broke, what worked. secrets stripped before anything leaves the machine.
1
1
27
and the agent doesn't need to remember to ask: a session-start hook feeds recent project context in automatically, on all six harnesses. shipped cursor support 1 day after someone opened the issue. more parsers welcome.
1
4
vshulcz/deja-vu: Memory layer for coding agents: search, MCP recall, auto-context, secret redaction, stats, share and sync over the session logs Claude Code, Codex and opencode already write. One zero-dep binary. #devopsish github.com/vshulcz/deja-vu/
1
2
148
vshulcz retweeted
Unlock your coding agent's memory with Deja Vu. See link below 👇
1
1
1
38
Wrote up how I got a Python DI container's resolve from ~53 µs/op down to 0.40 - almost hand-wiring speed. Plan caching, deleting a check that can't fire, and compiling the dependency graph with exec + CSE. vshulcz.hashnode.dev/python-…
1
28
Released Injex 1.3.0. Tiny typed dependency injection for Python apps: - zero runtime deps - constructor injection from type hints - singleton / transient / scoped lifetimes - test overrides - graph validation before startup - fast repeated resolves github.com/vshulcz/injex
1
27
I usually avoid DI containers in Python. But I kept hitting the same boring problem: manual wiring drifting across APIs, CLIs, workers, and tests. I wrote about the trade-off, and the small type-hint driven container I built around it. vshulcz.hashnode.dev/when-py…
1
25