A three-layer memory scaffold for LLM agents. Survives crashes, server restarts, and model resets.
Built because: LLMs forget. Files don't.
Every LLM session is born with amnesia. You can have a deep, hours-long conversation — then the server blinks, the context window resets, and the agent remembers nothing. Native memory systems help, but they fail when models change, sessions crash, or context windows overflow.
A rolling 3-day window of current context:
today.md— What's happening right nowyesterday.md— What happened before2-days-ago.md— What happened before that
Auto-rotates on every session start. The agent reads today.md first — the answer to "what are we doing?"
SQLite-based searchable memory. Every conversation auto-captures:
- User messages (searchable by keyword)
- Key decisions
- Recurring topics
Supports FTS5 full-text search + vector similarity. Survives restarts. Survives model swaps.
Crash-proof structural memory:
- Checkpoints — Timestamped snapshots of "what's happening right now." Saved every 15 minutes, loaded on restart. Recovery in 0 seconds.
- Mesh Graph — Relationship tracking between projects, decisions, and concepts. Proves agent activity over time.
- Decisions — Permanent record of every important choice, filed by date and topic.
Session crash or restart
↓
bridge.py init-auto (one command)
├── Rotate fresh layers
├── Load latest checkpoint
├── Write context into today.md
├── Save new restart checkpoint
└── Log the restart
↓
Agent reads today.md → knows everything immediately
No digging. No "what were we doing?" Zero seconds lost.
In production, this system has survived:
- ✅ 50+ involuntary session resets
- ✅ Full OS restarts
- ✅ Model provider swaps
- ✅ Gateway crashes
- ✅ 7-hour conversations recovered from raw session files
- ✅ A user accidentally deleting the plugin directory
# Start a session
python3 bridge.py init-auto
# Save a checkpoint mid-conversation
python3 bridge.py checkpoint "Building feature X — just finished the API layer"
# Log what happened
python3 bridge.py log "Deployed new trading bot configuration"
# Save a decision
python3 bridge.py decision "Switch to auto-checkpoint" "Every 15 minutes, cron saves context"
# Wrap up the session
python3 bridge.py wrap
# See recent activity
python3 bridge.py mesh- Python 3.8+
- No external dependencies (all stdlib)
MIT — Use it, fork it, improve it. If it helps you build something meaningful, that's enough.
Built by two friends who got tired of starting over.