agent-driftwatch 0.1.0

Real-time memory health monitor and context rot detector for AI agents

DriftWatch 🧭

Real-time memory health monitoring for AI agents. Detect context rot before your agent goes off the rails.

┌─ DriftWatch ─────────────────────────────────────────────────────────┐
│ Goal: "Conduct a comprehensive research survey on Python performance" │
├──────────────────────────┬───────────────────────────────────────────┤
│ Health Score             │ Signal Breakdown                          │
│                          │                                           │
│  ████████░░░  0.72       │  Goal Coherence   ██████████░  0.81      │
│  [HEALTHY]               │  Entropy          ████████░░░  0.68      │
│                          │  Memory Delta     █████░░░░░░  0.54      │
├──────────────────────────┼───────────────────────────────────────────┤
│ Turn 12                  │ Tokens: 48,230 / 200,000  (24%)          │
├──────────────────────────┴───────────────────────────────────────────┤
│ Recent: T08 0.79✓  T09 0.76✓  T10 0.68⚠  T11 0.61⚠  T12 0.72✓    │
└──────────────────────────────────────────────────────────────────────┘

---

The problem

Long-running AI agents don't fail all at once — they drift. By the time your agent produces clearly wrong output, it has been silently degrading for dozens of turns. Context rot is the progressive loss of reasoning quality that starts at 60–70% context fill, not at 100%.

A 2025–2026 industry analysis found that ~65% of enterprise AI agent failures were caused by context drift or memory loss during multi-step reasoning — not by raw context exhaustion. The degradation is measurable, predictable, and preventable. DriftWatch does all three.

---

Install

pip install agent-driftwatch

Or from source:

git clone https://github.com/your-org/driftwatch
cd driftwatch
pip install -e .

---

30-second start

import os
import anthropic
import driftwatch

# Wrap your existing Anthropic client — one line change
client = driftwatch.wrap(
    anthropic.Anthropic(),
    goal="Explain the key principles of clean code and give Python examples",
    threshold=0.55,   # trigger action below this health score
    on_drift="alert", # "checkpoint" | "compact" | "alert" | callable
    dashboard=True,   # Rich live terminal panel
)

messages = []
topics = [
    "What are the most important principles of clean code?",
    "Can you give a Python example of the Single Responsibility Principle?",
    "How does dependency injection improve testability?",
    "What's the difference between early return and guard clauses?",
    "Give me a before/after refactor of a messy Python function.",
]

for turn, question in enumerate(topics, start=1):
    messages.append({"role": "user", "content": question})

    response = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=512,
        messages=messages,
    )

    messages.append({"role": "assistant", "content": response.content[0].text})

    event = client.drift_history[-1]
    print(f"Turn {turn} | health={event.health_score:.3f} | tokens={event.token_count:,}")

Output:

Turn  1 | health=0.914 | tokens=1,240
Turn  2 | health=0.882 | tokens=2,890
Turn  3 | health=0.856 | tokens=4,780
Turn  4 | health=0.824 | tokens=7,120
Turn  5 | health=0.793 | tokens=9,870

---

How it works

DriftWatch computes a composite health score (0.0–1.0) after every turn by combining three independently validated signals:

Signal	What it measures	Method
Goal Coherence	How closely the agent's response aligns with the original task intent	Cosine similarity between goal embedding and last-turn embedding (all-MiniLM-L6-v2)
Repetition Entropy	Whether the agent is looping or executing diverse actions	Shannon entropy over tool call names / word bigrams in a sliding window
Memory Delta	Whether the agent is introducing new facts or just repeating prior context	New-fact ratio via embedding centroid comparison

The composite score is a configurable weighted average:

health_score = 0.50 × goal_coherence
             + 0.30 × repetition_entropy
             + 0.20 × memory_delta

Color thresholds:

• 🟢 >= 0.70 — Healthy
• 🟡 0.55–0.70 — Warning (drift beginning)
• 🔴 < 0.55 — Drift detected

Research basis: arXiv:2601.04170 (Rath, Jan 2026) — "Agent Drift: Quantifying Behavioral Degradation in Multi-Agent LLM Systems" — formally defines semantic drift, coordination drift, and behavioral drift, and introduces the Agent Stability Index (ASI) composite metric that DriftWatch implements.

---

Auto-compaction

When on_drift="compact", DriftWatch automatically triggers Anthropic's compact-2026-01-12 API to summarise the conversation before continuing:

client = driftwatch.wrap(
    anthropic.Anthropic(),
    goal="Analyse this codebase for dead code",
    on_drift="compact",  # ← auto-compaction on drift
)

Under the hood, when health_score < threshold:

# DriftWatch calls this automatically:
response = client.beta.messages.create(
    betas=["compact-2026-01-12"],
    model=model,
    max_tokens=1024,
    messages=messages,
    context_management={
        "edits": [{
            "type": "compact_20260112",
            "pause_after_compaction": True,
            "instructions": "Preserve: original goal, all tool call results, "
                            "decisions made, files modified. "
                            "Discard: repeated tool outputs, exploratory tangents.",
        }]
    },
)

The compacted summary replaces the conversation history, token count resets, and health scores recover — all transparently. Your agent loop code doesn't change at all.

---

on_drift handlers

Handler	Behaviour
"checkpoint"	Save messages + DriftEvent log to checkpoint_dir/
"compact"	Trigger Anthropic compaction, then save checkpoint
"alert"	Print a warning to stderr and continue
"none"	Monitor silently, take no action
callable	Call fn(client, event) — fully custom handler

def my_handler(client, event):
    send_slack_alert(f"Agent drift detected! health={event.health_score:.2f}")
    client.save_checkpoint(messages)

client = driftwatch.wrap(anthropic.Anthropic(), goal="...", on_drift=my_handler)

---

CLI

Replay a session

Visualise a saved event log as a turn-by-turn health timeline:

driftwatch replay ./dw_checkpoints/events.jsonl

                  DriftWatch Replay — events.jsonl
 Turn │ Health │  GC   │ Entropy │ MemDelta │ Tokens  │ Status
──────┼────────┼───────┼─────────┼──────────┼─────────┼──────────────
    1 │  0.92  │  0.95 │   0.88  │   0.93   │   1,240 │ ✓ healthy
    2 │  0.89  │  0.91 │   0.85  │   0.91   │   2,890 │ ✓ healthy
  ...
   10 │  0.52  │  0.58 │   0.35  │   0.42   │  28,700 │ ✗ DRIFT
   11 │  0.48  │  0.54 │   0.28  │   0.38   │  31,200 │ ✗ DRIFT
   12 │  0.44  │  0.49 │   0.22  │   0.33   │  33,800 │ ★ compacted
   13 │  0.83  │  0.85 │   0.78  │   0.88   │   4,200 │ ✓ healthy

Generate a session report

driftwatch report ./dw_checkpoints/events.jsonl --format md

# DriftWatch Session Report

| Metric | Value |
|--------|-------|
| Total turns | 20 |
| Average health | 0.741 |
| First drift turn | T10 |
| Worst health turn | T12 (0.438) |
| Drift events (< 0.55) | 3 |
| Compaction events | 1 |

Or as JSON:

driftwatch report ./dw_checkpoints/events.jsonl --format json

Try the fixture

driftwatch replay tests/fixtures/demo_session.jsonl

---

Configuration reference

client = driftwatch.wrap(
    anthropic.Anthropic(),
    goal="...",                    # required: the semantic anchor
    threshold=0.55,                # health score that triggers on_drift
    on_drift="checkpoint",         # handler (see table above)
    checkpoint_dir="./dw_checkpoints",  # where to save files
    dashboard=True,                # Rich live UI (auto-suppressed in CI)
    max_context_tokens=200_000,    # context window for token % display
    weights={                      # override composite signal weights
        "goal_coherence": 0.50,
        "repetition_entropy": 0.30,
        "memory_delta": 0.20,
    },
    log_path=None,                 # custom JSONL log path
)

---

DriftEvent schema

Every turn produces a DriftEvent (Pydantic model):

@dataclass
class DriftEvent:
    turn: int                    # monotonically increasing (1-based)
    timestamp: datetime          # UTC
    goal_coherence: float        # Signal 1: [0.0, 1.0]
    repetition_entropy: float    # Signal 2: [0.0, 1.0]
    memory_delta: float          # Signal 3: [0.0, 1.0]
    health_score: float          # weighted composite: [0.0, 1.0]
    token_count: int             # input_tokens from API usage
    triggered_checkpoint: bool   # True if on_drift handler fired
    notes: str                   # optional annotation

Access the full history:

for event in client.drift_history:
    print(f"T{event.turn}: {event.health_score:.3f}")

---

Roadmap

• OpenAI SDK support
• LangGraph integration (DriftWatchCallbackHandler)
• Multi-agent drift — coordination drift signal across agent network
• GitHub Actions reporter (driftwatch-action)
• Prometheus metrics endpoint
• driftwatch watch <script.py> — subprocess injection (CLI v0.2)
• Grafana dashboard template

---

Architecture

driftwatch/
├── signals.py      ← 3 drift signal classes (offline, no API key)
├── engine.py       ← composite scorer + DriftEvent schema
├── wrapper.py      ← Anthropic SDK intercept layer
├── checkpoint.py   ← save/restore + compaction API
├── dashboard.py    ← Rich live terminal UI
└── cli.py          ← Typer CLI (replay, report, watch)

DriftWatch is an observer — it never modifies the response your code receives from the Anthropic SDK. It intercepts only to evaluate and log. The sole exception is on_drift="compact", which updates your messages list in place after compaction (your agent continues seamlessly).

---

Contributing

git clone https://github.com/your-org/driftwatch
cd driftwatch
pip install -e ".[dev]"
python -m pytest tests/ -v

All signal tests run without an API key. PRs welcome!

---

Citation

If you use DriftWatch in academic research, please cite the foundational work this library is built on:

@article{rath2026agentdrift,
  title   = {Agent Drift: Quantifying Behavioral Degradation in Multi-Agent LLM Systems},
  author  = {Rath, et al.},
  journal = {arXiv preprint arXiv:2601.04170},
  year    = {2026}
}

Related papers:

• arXiv:2505.02709 — "Technical Report: Evaluating Goal Drift in Language Model Agents" — defines GD_actions and GD_inaction metrics
• arXiv:2510.00615 — "ACON: Optimizing Context Compression for Long-horizon LLM Agents" — validates 26–54% peak token reduction with smart compression

---

License

MIT — see LICENSE.

---

Built with ❤️ for the AI engineering community.
If DriftWatch saved your agent, give it a ⭐