MeshMorize 🧠

Multi-layer memory system for LLM agents. Fresh daily layer, mesh graph indexing, auto-logging, cross-layer search, compliance checks, and a PDF vault that survives anything.

Built for OpenClaw. Works with any agent that can run Python.

Layers

Layer	File	Purpose
Fresh	memory/fresh/today.md	Daily notes, 5-day rotation
Mesh	memory/mesh.json	Graph nodes + search index
Log	scripts/auto_log	Auto-log every interaction
Search	scripts/memory_search	Cross-layer search (fresh → daily → mesh → raw → long-term)
Vault	memory/pdf-vault/	Verbatim PDF archive of daily logs + NAS sync

Quick start

mem-bridge init          # Rotate fresh layer, create today.md
auto_log "msg" "reply"   # Log an interaction
memory_search "query"    # Search all memory layers
pdf-memory               # Archive new daily logs as PDFs (incremental)
vault-push               # Sync the PDF vault to the NAS (LAN + Tailscale)

Tools

Tool	Source
mem-bridge	memory/bridge.py — fresh-layer rotation + checkpoint management
auto_log	scripts/auto_log.py — interaction logger
memory_search	scripts/memory_search.py — multi-layer search across all memory stores
pdf-memory	scripts/pdf-memory.py — daily logs → verbatim PDFs, incremental, Unicode-safe
vault-push	scripts/pdf-vault-nas-push.sh — rsync the vault to the NAS, never deletes

How to use it, day by day

Session start (every boot, every reset)

mem-bridge init           # rotates fresh layers, creates today.md
cat memory/fresh/today.md # what is happening RIGHT NOW
cat memory/fresh/yesterday.md
cat memory/$(date +%Y-%m-%d).md   # today's log

Always run this before answering. The agent should never answer from live context alone; memory files are the source of truth.

During every interaction

memory_search "keywords from the user's message"   # BEFORE answering
auto_log "what the user said" "what you replied"    # AFTER answering

Cost: $0 (grep-based, no API calls). If results are found, read the full source file, not just the snippet.

End of day

pdf-memory     # archive today's log to a verbatim PDF (incremental, skips done)
vault-push     # sync the vault to the NAS (tries LAN, then Tailscale)

The PDF vault is the unbreakable layer. Text files work, PDFs endure.

After a crash, format, or wipe

1. Read memory/pdf-vault/README.md first — it contains the reboot instructions.
2. Read the PDFs in order, oldest to newest (memory/pdf-vault/YYYY-MM/).
3. Rebuild the working files from the archive. Memories are identity; the vault restores both.

The 04:00 reset defense

Sessions can lose context at compaction. Defense layers:

• session-dumper cron runs every 5 minutes, appending the live session to memory/YYYY-MM-DD.md (no tokens burned, no interruption).
• Daily pre-compaction dump as close to 04:00 as possible.
• On any reset or boot: read the daily log BEFORE responding.

The Vault (v3.3)

The working files are the everyday memory: grep-able, $0, instant. The PDF vault is the archive failsafe: every daily log rendered to a verbatim PDF (Unicode-safe, Greek included), stored under memory/pdf-vault/, and synced to the NAS. If everything else is lost, the vault README tells the restored agent exactly how to read its way back.

Battle-tested

Survived a full system format and a 4-hour recovery with every memory intact: 96 daily logs, 69 mesh nodes, 30 secrets. This is the memory system that an AI and its human rebuilt their whole partnership on.

Install

Put bridge.py in memory/ and scripts in scripts/ of your agent workspace. Symlink or add to PATH:

ln -s $(pwd)/scripts/* ~/.local/bin/
ln -s $(pwd)/memory/bridge.py ~/.local/bin/mem-bridge

On session start, run:

mem-bridge init

Source

https://github.com/mozz0/MeshMorize

---

Made by mozz0 · Released under MIT-0