context-pruning-directive

Dynamic Context Pruning (Sliding Memory Compactor)

Overview

In long-running autonomous agent sessions, every new message appends the entire historical transcript to the input array. If an agent executes 40 tool calls (reading files, executing bash commands, running test suites), historical tool outputs accumulate in the context window.

This unpruned history causes Quadratic Token Growth ($O(N^2)$): a 5,000-token file read at Turn 2 is needlessly re-sent on Turns 3 through 50, consuming 240,000 redundant input tokens and causing "Lost-in-the-Middle" Attention Degradation.

The Dynamic Context Pruning Protocol actively sanitizes historical turns: replacing stale, completed tool outputs with compact 1-line tombstone summaries and maintaining a high-density rolling memory buffer.

---

Unpruned Quadratic Accumulation vs. Dynamic Context Pruning

┌─────────────────────────────────────────────────────────────┐
│                 Context Growth Dynamics                     │
│                                                             │
│  Unpruned Transcript ($O(N^2)$ Quadratic Explosion):        │
│  • Turn 2: Reads 800-line file (4,000 tokens)               │
│  • Turn 3..50: 4,000 tokens re-sent on every turn           │
│  ↳ Total Waste: 4,000 tokens $\times$ 48 turns = 192,000 tokens! │
│  ↳ Latency increases from 1.2s to 18.5s per turn            │
│                                                             │
│  Dynamic Context Pruning ($O(N)$ Linear Bounded Memory):    │
│  • Turn 2: Reads 800-line file $\rightarrow$ Edits target lines    │
│  • Turn 3: Old read pruned $\rightarrow$ `[Read: auth.ts (lines 40-50)]`│
│  ↳ Context remains strictly bounded $< 6,000$ tokens        │
│  ↳ 78% Cost Reduction, Stable Sub-second Latency            │
└─────────────────────────────────────────────────────────────┘

---

The 3-Tier Context Pruning Strategy

┌───────────────────────────────────────────────────────────────────────────┐
│ 1. EPHEMERAL TOOL PRUNING: Compress old tool outputs older than 2 turns   │
│    `cat package.json` (300 lines) $\rightarrow$ `[Tool Output: package.json verified]`│
│                                                                           │
│ 2. SLIDING CONVERSATION WINDOW: Keep only last $K=6$ turns verbatim       │
│                                                                           │
│ 3. MILESTONE STATE CONDENSATION: Summarize completed phases into 3 bullets│
└───────────────────────────────────────────────────────────────────────────┘