forge-guardrails 0.1.2

Foundation types for an LLM-agent workflow framework
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# Context & Memory Module

This module is responsible for monitoring token consumption, warning about context window limits, and performing transcript compaction when approaching token budgets.

## File Breakdown

- [hardware.rs]file:///Users/whit3rabbit/Documents/GitHub/forge-rs/src/context/hardware.rs: Probes system hardware configurations to determine native memory bounds. Supports:
  - Apple Silicon unified memory extraction via `system_profiler` plist parsing.
  - NVIDIA VRAM profiling via `nvidia-smi` parsing.
- [manager.rs]file:///Users/whit3rabbit/Documents/GitHub/forge-rs/src/context/manager.rs: The central [ContextManager]file:///Users/whit3rabbit/Documents/GitHub/forge-rs/src/context/manager.rs#L44 which manages active token counts, tracks the history, triggers warnings, and executes compaction routines.
- [strategies.rs]file:///Users/whit3rabbit/Documents/GitHub/forge-rs/src/context/strategies.rs: Contains concrete algorithms for compacting chat history:
  - `NoCompact`: A basic strategy that raises errors when the budget is exceeded.
  - `SlidingWindowCompact`: Drops oldest messages while preserving initial system setup messages.
  - `TieredCompact`: Multi-phased compaction that replaces detailed tool calls and responses with compacted summary logs to fit budgets.
- [mod.rs]file:///Users/whit3rabbit/Documents/GitHub/forge-rs/src/context/mod.rs: Module exports and base helper utilities.