Deja Vu works best with agents that can:
- read repo files
- follow project instructions
- update files
- respect memory budgets
- avoid storing secrets or noisy transcripts
Any capable coding agent can follow the protocol. The memory lives in the repo, not inside one vendor.
Use the normal Deja Vu flow when an agent can inspect and edit project files:
- read
AGENTS.md - scan
memory/impressions.jsonl - load
memory/summary.mdonly for weak familiarity - load 1-3 linked detailed records only for strong familiarity
- write back durable memory only
This is the intended path for file-aware coding agents such as Codex, Claude Code, Cursor, Windsurf, Gemini CLI, and similar tools when they can access the repo.
Some chat agents cannot directly read repo files.
Use the manual path:
- paste
AGENTS.md - paste
memory/impressions.jsonlor the relevant lines - paste
memory/summary.mdonly when the task needs project context - paste 1-3 detailed records only when the cue strongly matches
- ask the agent to propose durable writeback, then review before committing it
Do not paste the whole memory tree by default.
Say:
Any capable coding agent can follow the protocol.
Do not say:
Every agent automatically supports Deja Vu.
Deja Vu is a repo-local convention and memory discipline. Agents still need enough capability and instruction-following reliability to use it.
Short prompts are included in starter-kit/prompts:
- Codex
- Claude Code
- Cursor
- Windsurf
- ChatGPT
- Gemini CLI
They are intentionally similar. The protocol should stay stable even when the tool changes.