> ## Documentation Index
> Fetch the complete documentation index at: https://docs.compose.market/llms.txt
> Use this file to discover all available pages before exploring further.

# Minimal Token Prompting

> How Manowar keeps agent prompts compact without hiding identity, memory, or session context.

Manowar uses a dynamic system block built per turn in `framework.ts`. The block is small on cold starts and grows only when memory or session context has useful data.

## Prompt Sections

| Section             | Source                                         | Included when                                            |
| ------------------- | ---------------------------------------------- | -------------------------------------------------------- |
| Identity            | `harness/identity.ts` or agent config fallback | Always.                                                  |
| Persona             | Registered `systemPrompt`                      | Present on the agent config.                             |
| Memory              | `retrieveAgentMemory()`                        | Ranked memory exists for the user turn.                  |
| Session             | Per-request execution context                  | Session budget, permissions, or Backpack accounts exist. |
| Capability sentence | `harness/tools/surface.ts`                     | Always.                                                  |

## Memory Budget

The runtime retrieves all six layers but injects only compact results:

| Setting            | Value                                                          |
| ------------------ | -------------------------------------------------------------- |
| Layers             | `working`, `scene`, `graph`, `patterns`, `archives`, `vectors` |
| Retrieval limit    | `12`                                                           |
| Max injected items | `6`                                                            |
| Max chars per item | `120`                                                          |
| Prompt budget      | `900` chars                                                    |

The broader memory summary formatter supports a larger 1,800 character budget for non-turn uses. The agent turn path uses the tighter 900 character budget.

## Compacting Rules

Memory is deduped by fingerprint, ranked by layer priority, semantic score, query overlap, confidence, recency, and access count, then packed by score-per-character. This keeps the injected context short without throwing away the layer signal.

## Comparison

| System            | Context strategy                                                         | Manowar strategy                                                                        |
| ----------------- | ------------------------------------------------------------------------ | --------------------------------------------------------------------------------------- |
| LangGraph         | State is graph-defined and persisted through checkpointers/stores.       | The native turn prompt receives compact ranked context from the memory system.          |
| Google ADK        | Session, State, and Memory manage conversational context.                | Manowar scopes memory by agent wallet, user, thread, workflow, and local HAI id.        |
| OpenAI Agents SDK | The app supplies instructions, tools, handoffs, guardrails, and context. | The runtime assembles identity, memory, session, and tools for deployed Compose agents. |

## Related

* [Memory introduction](/manowar/memory/introduction)
* [Lazy discovery](/manowar/lazy-discovery)
* [Agent-to-agent](/manowar/agent-to-agent)