Agent Session Persistence — 118 sessions of crash test data

The Problem

Claude Code has no session persistence. Context fills, conversation dies, all state is lost. Autonomous agent loops are impossible without external orchestration.

I've run 118 autonomous Claude Code sessions building a startup. Not conversations — sessions with handoff protocols, context lifecycle management, multi-memory persistence, and autonomous cycling that runs overnight without human input.

Your product doesn't support any of this natively. So I built it. Here's what I learned.

What I Built (118 Sessions, 4 Iterations)

Context Guardian — PreToolUse hook that scores every tool call, tracks cumulative context consumption, escalates through reminder levels, then blocks ALL tools except the shutdown command. Prevents the agent from working past its cognitive cliff.

Sleep Cycle Protocol — Deterministic 3-phase shutdown: (1) persist state to multi-memory + write handoff contract, (2) write autonomous work queue, (3) external VBS bridge creates fresh conversation and restarts. v4.1 includes mutex locking for multi-session overlap prevention.

Multi-Memory Architecture:

• Structured notes (MCP server) — decisions, summaries, searchable by title/tag
• Knowledge graph (MCP server) — entities + relations, cross-session references
• Semantic search (custom BM25 + vector hybrid) — 2,135 chunks, 222 sources
• Tiered query protocol: all 3 stores queried in parallel before any file read

Handoff Protocol: 6-field contract written at every session close. Next session picks up in 5 seconds. Zero ambiguity. Fields: completed, stopped-at, next-action, decisions, open-threads, files-modified.

Session Type Routing: 4 types (Sprint/Decision/Pipeline/Orientation) with different context budgets (60%/40%/80%/35%), different tool access rules, and different close conditions. Hook-enforced via temp file state.

Autonomous Work Queue: JSON marker file read at session start. If present → auto-enter Sprint mode, execute queue items sequentially, sleep at threshold, cycle repeats. No human input required. Runs overnight.

10 Real Failure Modes (of 40+)

| # | What Broke | How I Fixed It |
|---|-----------|---------------|
| 1 | /clear kills webview DOM — external input dies | newConversation (Ctrl+Alt+N) preserves webview |
| 2 | Clipboard paste fails in VS Code webview | SendKeys types directly into intact webview |
| 3 | Multiple VBS scripts race-condition on overlap | Mutex lock file + taskkill pre-launch |
| 4 | Compound knowledge graph queries return empty | Decompose to parallel single-concept queries |
| 5 | Guardian can't force shutdown, only nags | Block ALL tools except /sleep at critical threshold |
| 6 | Guardian state persists across new conversations | Reset temp state file at session start |
| 7 | Deleting autonomous marker triggers permission prompt | Read-only — overwrite on next cycle, never delete |
| 8 | Single focus attempt misses panel after new conversation | 3x focus attempts with 1.5s gaps |
| 9 | Retry attempts stomp on already-running session | Single attempt after generous wait |
| 10 | Session startup burns 30%+ context before real work | Session type routing with budget allocation |

What Native Implementation Needs

1. Session persistence API — save/restore structured state across context resets. Not just chat history — decisions, work queues, behavioral rules.
2. Context lifecycle hooks — native cognitive load estimation + graceful shutdown trigger. Not a token counter — a cognitive load estimator.
3. Multi-memory tiers — session (short-term), project (medium-term), user (long-term). Different query patterns for each.
4. Autonomous agent loops — execute → save state → reset → resume at API level. No VBS. No keyboard automation.
5. Session type contracts — declarative mode switching with tool access rules and context budgets.
6. Handoff protocol standard — formalize the minimum viable state contract between sessions.

About Me

I'm a BCG consultant building a startup entirely inside Claude Code. I'm not a developer tools person — I'm a power user who solved infrastructure problems because the native tools didn't exist yet. That's exactly the perspective your team needs.

Happy to walk through the full architecture, share the codebase, or advise on the native implementation. 118 sessions of crash test data is yours if you want it.

Contact: jsaksham05@gmail.com | GitHub: @sakshamking