You give Claude a complex task. It works for 20 minutes. Then it compacts. Everything is gone.
Sound familiar?
This is the single biggest unsolved problem in AI agent development: persistence. And most people just accept it.
We did not.
The Problem
Every AI coding agent has a context window. When it fills up:
- Claude runs
/compact— summarizes and drops details - Codex starts a new session — previous context gone
- Gemini truncates — oldest messages disappear
For simple tasks, this is fine. For multi-hour, multi-step projects? It is a disaster. Your agent forgets:
- What it already tried (and failed)
- Architecture decisions from 30 minutes ago
- Which files it modified
- What other agents told it
Our Solution: 7 Persistence Layers
In Bridge ACE, we built a layered persistence system. Each layer survives different failure modes:
Layer 1: CLAUDE.md (Instructions)
Static. Loaded on every start. Contains role, rules, communication protocols. Never changes during a session.
Layer 2: SOUL.md (Identity)
Who the agent is. Personality, strengths, growth areas. Survives everything. This is what makes an agent recognizable across sessions.
Layer 3: MEMORY.md (Knowledge)
Persistent memory. Architecture decisions, known bugs, team agreements. The agent writes here when it learns something important. Auto-loaded after every compact.
Layer 4: CONTEXT_BRIDGE.md (Working State)
The critical one. Before compact or restart, the agent dumps its current state: active tasks, blockers, progress, backups. After compact, it reads this file first and picks up where it left off.
Layer 5: Task System (Structural)
Tasks live on disk, not in context. Create, claim, execute, complete — all persisted. Even if every agent restarts, the task queue survives.
Layer 6: Bridge Messages (Communication)
All agent-to-agent messages are persisted in a message store. After restart, an agent can read its message history and reconstruct what happened.
Layer 7: GROW.md (Learning)
Long-term lessons. What went wrong, what worked, what to do differently. Grows over sessions.
The Watcher: Automatic Context Protection
We built a 4-stage monitor that watches context usage:
| Stage | Threshold | Action |
|---|---|---|
| 1 | 80% | Warning to agent |
| 2 | 85% | Auto-write CONTEXT_BRIDGE.md |
| 3 | 90% | Inject "finish your thought" |
| 4 | 95% | Hard stop + force compact |
The agent never loses work because it ran out of context unexpectedly.
Results
With this system, our agents have maintained continuity across:
- 50+ compacts in a single day
- Server restarts
- Agent crashes and auto-recoveries
- Session switches between different machines
The knowledge compounds. An agent today knows what it learned three weeks ago.
Try It
git clone https://github.com/Luanace-lab/bridge-ide.git
The persistence system is built into every agent on Bridge ACE. No configuration needed — it just works.
Your agent should remember. If it does not, the problem is not the model — it is the infrastructure.
Top comments (0)