The Problem
Every agent builder hits the same wall. You pile on context. The agent still forgets. The instinct is to increase the context window - make the desk wider. But a bigger desk doesn't give the agent a filing cabinet, and it doesn't give it a recording of the meeting it just sat through.
“Me: Remember last time with Acme? We decided not to discount below X.
Agent: Which customer is Acme? Which decision?”
It read every word. It just didn't have anywhere to look it up.
The Two Patterns
GBrain - Cross-Conversation Knowledge
A query/index layer on top of your company wiki. When a new hire shows up on day one, you don't make them read every Slack message. You hand them a Notion link. GBrain does the same for agents - it indexes your markdown knowledge base so agents can query people, projects, and decisions before acting.
- • Use when: agent works across conversations, across tasks, across team members
- • Common thread: “between conversations”
- • github.com/garrytan/gbrain
Lossless Claw - Within-Conversation Memory
Like a full meeting recording. The raw messages of the current conversation are kept intact. When the model auto-compresses history, the original text stays searchable and recoverable. Nothing is lost - summaries link back to source messages in a DAG.
- • Use when: long conversations (50+ turns), after auto-compression, debugging agent decisions
- • Common thread: “within a conversation”
- • github.com/Martian-Engineering/lossless-claw
Why This Matters for Agent Security
Both patterns are critical for safe agent operations - and both are currently missing from most agent deployments. When agents can't recall past decisions, they repeat mistakes. When they can't verify facts before acting, they take unsafe actions. When conversations get compressed without recovery, audit trails disappear.
This is the memory gap that complements the governance gap. AgentVet addresses the question: “Is this agent safe to run?” GBrain and Lossless address: “Can this agent remember enough to act safely?”
The 5-Layer Diagnostic
Next time your agent “forgets,” run through these layers in order:
- Capture - Did the fact ever enter any system?
- Lossless - Did the conversation survive, or did compression eat it?
- GBrain - Can it be fetched across conversations?
- Ranking - Did the right fact surface before acting?
- Task - Did the current task tell the agent why this fact matters?
Most “forgetting" bugs live in layer 3 or 4, not 1 or 2.
References
- GBrain - Shared brain layer. Indexes markdown knowledge across sessions.
- lossless-claw - OpenClaw context-engine plugin. DAG-based summarization with full recovery.
- OpenClaw - Agent runtime. The platform both tools plug into.
- Original article by VoxYZ - “Your Agent Needs a Wiki and a Recording, Not a Bigger Desk”
About AgentVet
AgentVet is building the governance layer for the AI agent ecosystem. We vet, scan, and certify agent skills so businesses can deploy agents safely. Learn more →