Open-source memory for coding agents, synced over SSH
TL;DR Highlight
Indexing on-disk session logs lets coding agents recall fixes they already found
Who Should Read
A developer switching between Claude Code, Codex and Cursor who is tired of agents re-debugging the same bugs
Core Mechanics
- It indexes session logs already on disk
- It reads sessions saved by 42 agents including Claude Code, Codex and Cursor.
- Months of pre-install history is searchable on day one.
- It runs locally with no LLM
- One Go binary, no model and no server.
- Embeddings are used only if you enable semantic recall.
- Agents pull memory without being asked
- Recall arrives at session start, before edits, and after failed commands.
- One MCP tool costs 477 definition tokens per turn versus 8,283 for the largest of seven servers measured.
- Recall and speed numbers are published
- 97.2% R@5 on LongMemEval-S, versus 96.6% MemPalace and 95.2% agentmemory as published.
- Indexing 19,195 sessions takes 17.6s; the next tool takes 72s.
- It fills gaps left by compaction
- Across 43 measured compactions the summary kept 77% of decisions and 0.2% of commands.
Community Reactions
- One commenter reviewed 140+ memory systems automatically
- They shared comparisons and a deja-vu review.
- Many people built their own, commenters said
- One uses bge-en-base CPU embeddings, an S3-like endpoint and SQLite, and it works well.
- They add a periodic consolidation step.
- A counterview says a curated MEMORY.md is enough
- Short, readable and easy to fix, versus overengineering.
- Others complained agents don't even read saved memories.
- The author shared their own numbers
- About 3.3GB of history, ~10s first index, 7-9ms warm searches.
- A headless Mac mini works overnight and the laptop pulls its memory in the morning.
How to Apply
- Install and ask about an old fix
- Run the install script and deja install --auto on macOS/Linux.
- In a new session, ask about something solved months ago and check that recall appears.
- Connect machines with sync ssh
- Run deja sync ssh <host> to exchange memory.
- Run deja secrets first to check source transcripts for credentials.
- Surface repeat failures with deja friction
- List errors that hit three or more sessions.
- Put the fixes into deja rules to propagate to every agent.
Try It
# Paste in terminal (macOS/Linux)
curl -fsSL https://raw.githubusercontent.com/vshulcz/deja-vu/main/install.sh | sh
deja install --auto
# Try searching
deja "jwt refresh token"
deja friction
deja secrets
# Next: sync with another machine
deja sync ssh {{host}}Terminology
Related Papers
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.
Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.
Harness Engineering 101: The LLM is the brain, the harness is the body
An agent is a loop managing a JSON array; 16 chapters build one from scratch
Harness2 by Google Research – Recursive Agent Harnessing
Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.
Ousted Savannah police officer gave ICE agent husband access to Flock system
A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
Running Claude Code agents in tmux works well, but limits appear around four agents.