Context Management & Compression

As agents do more work per conversation, managing the context window becomes critical. Portal One includes a three-layer context management system that keeps agents effective across long, complex conversations.

Token Accounting

Every model has a context window limit. Portal One tracks token usage per model and per conversation so agents know exactly how much context budget remains.

• Token counts are tracked for all messages, tool calls, and system prompts
• Budgets are model-specific (GPT-4 has different limits than Claude)
• Agents can see their remaining budget and plan accordingly

Tool Output Compression

Tool calls (especially terminal output, file reads, and web extracts) can return large amounts of data. Portal One applies two compression strategies:

Eager Truncation

Large tool outputs are truncated at the time of return. If a terminal command produces 10,000 lines of output, the agent sees a manageable summary with the most relevant parts preserved.

Retroactive Compression

As the conversation grows, older tool results are compressed further. A detailed code diff from 20 messages ago gets summarized to its key findings, freeing up context budget for the current task.

History Compression

When a conversation approaches its context limit, Portal One applies intelligent history compression:

• Older messages are summarized rather than dropped entirely
• Key facts and decisions are preserved in the summary
• Recent messages remain at full fidelity
• The agent maintains continuity without losing important context

The Result

Context management is fully automatic. You don't need to configure anything — agents stay sharp across long conversations, complex multi-step tasks, and extended back-and-forth interactions without hitting token limits or losing track of what's been discussed.