How Claude Code Achieves Unlimited Context with Multi‑Layer Caching and Self‑Evolving Agents

Two‑Layer System Prompt Cache

Claude Code splits the system prompt into a static global cache (shared across users) and a dynamic per‑session cache. The static part contains identity, rules, tasks, actions, tool guides, tone, and efficiency, while the dynamic part stores session‑specific guidance, memory, environment info, MCP instructions, language, output style, scratchpad, token budget, and boundary markers.

Cache Management Mechanism

Static cache is immutable and shared; dynamic cache is cleared only on /clear or /compact. The boundary marker separates the two, ensuring that changes in the dynamic part never invalidate the static cache.

Four‑Stage Compact Architecture

Claude Code applies a progressive compression strategy when the context limit is approached:

Micro‑Compact: incremental deletion of tool results using cache_edits.

Session‑Memory Compact: keep the most recent 5‑40 KB of messages and replace older conversation with a session memory block.

Full Compact: fork an agent to summarize the whole dialogue, reuse the prompt cache, and replace images with placeholders.

PTL Retry: if compression still exceeds limits, truncate the oldest message groups up to 20 % of the conversation, retrying up to three times.

Proactive Mode

When PROACTIVE or KAIROS is enabled, Claude Code switches to a completely different system prompt that gives the agent autonomy. Features include autonomous identity, timed wake‑ups, background task management, focus awareness, and sleep scheduling. The agent behaves independently when the terminal is unfocused and switches to collaborative mode when the user returns.

Token‑Budget Driven Work

Users can specify a token budget (e.g., “+500k”). The agent works until it reaches 90 % of the budget, stopping early if diminishing returns are detected for three consecutive rounds.

Skill Discovery

The system automatically scans the .claude/skills/ directory each turn (auto discovery) and can actively request additional skills (DiscoverSkillsTool). Remote skills are loaded on demand after being discovered.

Multi‑Agent Collaboration

Claude Code supports three collaboration models:

Fork: a background agent inherits the full prompt cache, performs heavy work, and returns only the final result.

Subagent: specialized expert agents (explore, verification, custom) run with independent prompts and tool sets.

Swarm: a persistent team of agents communicates via a file‑based mailbox, with a leader handling all permission dialogs.

Remote/Bridge Architecture

The agent runs in the cloud and can be accessed from any client (CLI, VS Code, web). Two transport stacks are available: WebSocket + HTTP POST (v1) and SSE + CCRClient (v2). Up to 32 concurrent sessions are isolated using Git worktrees.

Enterprise‑Grade Security

A 20‑plus‑step Bash security classifier checks for dangerous characters, command substitution, unsafe variables, brace expansion, backslashes, quoting issues, control characters, jq system calls, and /proc access. Zsh‑specific checks protect against module loading, emulate, zpty, ztcp, and forced removal. Additional layers include sandbox execution, fine‑grained file permissions, network isolation, commit attribution, and cyber‑risk directives.

Telemetry and Privacy

Claude Code sends three telemetry streams: an API header (always sent), first‑party event logging to BigQuery, and optional Datadog events. Data includes core metadata, environment context, process metrics, user identifiers, and PII‑protected fields. Users can disable most telemetry with DISABLE_TELEMETRY=1 or other environment flags, but the API header remains embedded in the system prompt.

Design Philosophy

The architecture is built around cache awareness (static + dynamic layers, cache‑edits, boundary markers) and infrastructure awareness (TTL‑driven sleep, token‑budget control, focus detection). This combination enables unlimited context, self‑evolving memory, autonomous operation, and industrial‑scale reliability.

Takeaways for OpenClaw

OpenClaw can adopt Claude Code’s layered cache, cache‑edits mechanism, focus‑aware autonomy, and distributed permission model while preserving its own 9‑layer composability.