TERX — browser agent memory

// real benchmark — Groq API, openai/gpt-oss-120b, measured

173.9x

average speedup on warm runs

100%

token savings on cache hits

10/10

cache hit rate across task suite

task	llm steps	agent (cold)	terx (warm)	speedup	tokens	cost saved	hit
User Login Flow	4	3.05s · $0.0065	0.090s · $0.000	34.0x	1,985	100%	✓
Search and Filter	4	17.82s · $0.0108	0.099s · $0.000	179.3x	2,634	100%	✓
Multi-step Signup	6	41.05s · $0.0142	0.103s · $0.000	399.4x	4,339	100%	✓
E-commerce Product	4	9.84s · $0.0123	0.100s · $0.000	98.1x	2,798	100%	✓
Settings Toggles	4	15.42s · $0.0082	0.100s · $0.000	154.5x	2,131	100%	✓
Data Table Pagination	3	11.27s · $0.0093	0.088s · $0.000	128.7x	1,756	100%	✓
Support Ticket	4	18.35s · $0.0078	0.100s · $0.000	184.2x	2,130	100%	✓
Fuzzy Search	3	10.41s · $0.0055	0.083s · $0.000	124.8x	1,393	100%	✓
Profile Update	4	17.20s · $0.0104	0.077s · $0.000	223.7x	2,544	100%	✓
Complex Form	4	16.51s · $0.0074	0.086s · $0.000	192.2x	2,072	100%	✓
total / average	—	160.93s · $0.0925	0.926s · $0.000	173.9x	23,782	100%	10/10

Reproduce: GROQ_API_KEY=... python -m terx.benchmarks.real_agent

No-key proof: terx demo and terx eval-local

// quickstart

MCP server

Python library

Run benchmark

Works with Claude Desktop, Cursor, Windsurf

Start Chrome with debugging enabled:

# close all Chrome windows first
google-chrome --remote-debugging-port=9222 --no-first-run

Start TERX:

pip install terx
terx-server

Add to your MCP config:

// claude_desktop_config.json
{
  "mcpServers": {
    "terx": { "command": "terx-server" }
  }
}

Wrap repeatable workflows with browser_task_start and browser_task_finish. Pass variables and postconditions for safe replay. First run records. Matching repeats replay without another LLM call.

Wrap your existing agent — 3 lines

from terx.cdp.session import BrowserSession
from terx.cache.cache import MemoryCache, session_for

cache = MemoryCache()

	async with BrowserSession() as session:
	    bridge = session.bridge()

	    async with session_for(
	        cache, bridge, "login to salesforce",
	        variables={"email": "user@example.com", "password": "..."},
	        postcondition={"text_contains": "Welcome"},
	    ) as ctx:
	        if ctx.hit:
	            await ctx.replay()          # 0 tokens, ~80ms
	        else:
            await your_agent.run(...)   # agent runs, TERX records

Run the real LLM benchmark yourself

# free key at console.groq.com
	pip install "terx[benchmark]"
	cp .env.example .env   # fill in GROQ_API_KEY
	python -m terx.benchmarks.real_agent

Runs 10 real tasks. Prints the table above with your measured numbers.

# local Chrome demo + deterministic eval suite
	terx demo
	terx eval-local

// how it works

  ┌─────────────────────────────────────────────────┐
  │                  your agent                      │
  │          (browser-use / Claude / anything)        │
  └───────────────────────┬─────────────────────────┘
                          │
  ┌───────────────────────▼─────────────────────────┐
  │              TERX session_for()                │
  │  on miss → records CDP commands                  │
  │  on hit  → replays them directly                 │
  └───────────────────────┬─────────────────────────┘
                          │ raw WebSocket
  ┌───────────────────────▼─────────────────────────┐
  │                  Chrome                          │
  │           (remote debugging port 9222)            │
  └─────────────────────────────────────────────────┘

CDP Bridge

Raw asyncio WebSocket to Chrome. No Playwright subprocess. <50ms startup, ~2MB RAM.

DOM Extractor

Reads Chrome's Accessibility Tree (not raw HTML). Computes a fuzzy structural hash — survives CSS changes and A/B tests without breaking cache hits.

Muscle Memory Cache

SQLite. On success: stores CDP command sequence keyed by (domain, dom_hash, task). On future runs: replays directly. INSERT OR IGNORE — first successful recording is canonical.

Dynamic node ID translation

Chrome assigns new backendNodeIds each session. TERX re-snapshots on replay and maps old IDs to current equivalents by role + label matching.

// mcp tools

tool	what it does
`browser_task_start`	Start a cacheable workflow; replays immediately on cache hit
`browser_task_finish`	Store a successful workflow or discard a failed one; postconditions prevent bad hits
`browser_get_state`	AX tree snapshot — stable element IDs, no hallucination-prone HTML
`browser_navigate`	Navigate to URL (scheme-validated, blocks javascript: data: file:)
`browser_click`	Click element by stable ID using a remappable node handle
`browser_click_at`	Fallback physical click at an element center point
`browser_type`	Focus an input by stable ID and insert text
`browser_screenshot`	Returns hash ref, not base64 — no context window poisoning
`browser_screenshot_get`	Retrieve a screenshot by hash ref when image bytes are needed
`browser_scroll`	Scroll up/down
`browser_new_tab`	Open new tab
`cache_stats`	Hit rate, total savings, unique domains cached
`cache_invalidate`	Clear cache for a domain when the UI ships a redesign

// memory layer for browser agents