diff --git a/WORK.md b/WORK.md
new file mode 100644
index 0000000..a9a02e6
--- /dev/null
+++ b/WORK.md
@@ -0,0 +1,78 @@
+# What’s going on inside `mathstream`
+
+Figured I’d write down how the whole thing is wired, more like a brain dump than polished docs. If you’re diving in to fix something, this should save you a bunch of spelunking.
+
+## Directory map (so you know where to poke)
+
+```
+mathstream/
+  __init__.py    # re-export central
+  engine.py      # arithmetic guts
+  exceptions.py  # custom errors
+  number.py      # StreamNumber, manual GC, watcher
+  utils.py       # sqlite junk drawer
+
+test.py          # smoke/integration script
+```
+
+## StreamNumber – the heart of it
+
+`mathstream/number.py` owns the `StreamNumber` class. The class does a couple jobs:
+
+- wraps a file of digits (either you give us a path or a literal; literals get canonicalised and dropped into `instance/log/literal_<hash>.txt`).
+- streaming happens via `.stream(chunk_size)` so we never load the whole thing; every time we read we call `touch_log_file` so the usage timestamp keeps moving.
+- when a new stream gets created we check if it lives under `LOG_DIR`. If yes, we register it with the sqlite tracker (`register_log_file`) and also bump a ref counter via `register_reference`.
+- there’s a weakref finaliser plus a global `_ACTIVE_COUNTER` that keeps tabs on python-side references. If the object falls out of scope we run `_finalize_instance`, which decrements the counter and if it was the last one we call `release_reference` (that may nuke the file instantly).
+- explicit `free()` exists for people who want deterministic cleanup. It’s basically like `free()` in C: drop ref count and optionally delete the file now. There’s an alias `free_stream` and the class is a context manager so `with StreamNumber(...) as sn:` cleans up automatically.
+
+So any time you’ve got a staged result hanging around in memory, the watcher knows about it. Once you ditch it—either by `free()` or just letting the object die—the sqlite ref count drops.
+
+## Engine – maths without ints
+
+Living in `mathstream/engine.py`. All the operators (`add/sub/mul/div/mod/pow`) pull chunks from the `StreamNumber` inputs, normalise them into sign + digit strings, run grade-school algorithms, then write the result back into `LOG_DIR`.
+
+- `_write_result` is the important bit: writes to disk, calls `register_log_file`, then wraps the file in a new `StreamNumber`. Because of that call, every staged result is tracked automatically.
+- We’re careful about signs: division and modulo follow Python’s floor division rules. Divide-by-zero is intercepted and converted into `DivideByZeroError`.
+- `clear_logs()` wipes the folder and calls `wipe_log_records()` to empty sqlite so the next run isn’t polluted.
+
+## Exceptions
+
+`mathstream/exceptions.py` just defines `MathStreamError` and the more specific `DivideByZeroError`. Nothing fancy, just so we don’t leak raw `ZeroDivisionError`.
+
+## SQLite watcher (`mathstream/utils.py`)
+
+This is the garbage-collection HQ. On import we run `_ensure_db(reset=True)` so every run starts from a clean DB (no migrations, no surprises). Two tables:
+
+- `logs` → metadata about every staged file: created time, last access, access count.
+- `refs` → current reference count (think “how many StreamNumber instances think they own this file”).
+
+Important functions:
+
+- `register_log_file(path)` – ensure both tables have a row (initial ref count 0).
+- `register_reference(path)` – increments the ref count, updates last access, access count etc. Called whenever a new `StreamNumber` points at the staged file.
+- `touch_log_file(path)` – called from `.stream()` so we know the file is being read.
+- `release_reference(path, delete_file=True)` – the inverse of register. If the count hits zero we remove the DB row and (optionally) delete the file right away.
+- `collect_garbage(score_threshold)` – this is the periodic sweeper. Computes `score = age / ((ref_count + 1) * (access_count + 1))`. Bigger score means older + less used. If score >= threshold it gets unlinked and removed from DB. Negative thresholds blow up on purpose.
+- `tracked_files()` – dumb helper that dumps `{path: ref_count}` out of the DB.
+- `wipe_log_records()` – nukes both tables; used by `clear_logs`.
+
+## How cleanup flows
+
+1. You run an operation (`add`, `mul`, whatever). Result file lands in `LOG_DIR`, gets registered, comes back as a `StreamNumber`.
+2. You stream it or create more streams from it – metadata keeps getting updated via `touch_log_file`/`register_reference`.
+3. When you’re done, call `.free()` or just drop references. Manual free is immediate. Otherwise the weakref finaliser catches it eventually.
+4. `release_reference` is what actually removes the sqlite entries and unlinks the data file when there are no logical references left.
+5. If you still have detritus (e.g. you crashed before refs hit zero), run `collect_garbage(threshold)` to sweep anything whose age outweighs usage.
+6. `active_streams()` reports what’s still alive in Python land; `tracked_files()` shows what the DB thinks is referenced.
+
+## Example run (`test.py`)
+
+`test.py` is half regression, half reference script. It:
+
+- seeds some numbers, runs every operation, checks results.
+- makes sure `DivideByZeroError` fires.
+- frees every staged number to prove files vanish on the spot.
+- runs `collect_garbage(0)` just to make sure nothing else lingers.
+- dumps `active_streams()` and `tracked_files()` so you can see python vs sqlite state.
+
+If the logs ever seem suspicious, run that script—it’ll tell you immediately whether something’s still referenced or if the GC is forgetting to clean up.
diff --git a/mathstream/README.md b/mathstream/README.md
index f02ad9f..864044c 100644
--- a/mathstream/README.md
+++ b/mathstream/README.md
@@ -23,6 +23,7 @@ from mathstream import (
     pow,
     is_even,
     is_odd,
+    free_stream,
     collect_garbage,
 )
 
@@ -39,6 +40,9 @@ print("power =", "".join(pow(a, e).stream()))
 print("a is even?", is_even(a))
 print("b is odd?", is_odd(b))
 
+# drop staged artifacts immediately when you are done
+free_stream(b)
+
 # reclaim space for files whose age outweighs their use
 collect_garbage(0.5)
 ```
@@ -52,8 +56,9 @@ Each arithmetic call writes its result back into `instance/log` (configurable vi
 - **Automatic staging** – Outputs are stored under `LOG_DIR` with hashes based on input file paths, letting you compose operations without manual bookkeeping.
 - **Sign-aware** – Addition, subtraction, multiplication, division (`//` behavior), modulo, and exponentiation (non-negative exponents) all respect operand sign. Division/modulo follow Python’s floor-division rules.
 - **Utilities** – `clear_logs()` wipes prior staged results so you can start fresh.
+- **Manual freeing** – Call `stream.free()` (or `free_stream(stream)`) once you are done with a staged number to release its reference immediately. Logger metadata keeps per-path reference counts so the final free removes the backing file on the spot.
 - **Parity helpers** – `is_even` and `is_odd` inspect the streamed digits without materializing the integer.
-- **Garbage collection** – `collect_garbage(score_threshold)` computes a score from file age, access count, and reference count (tracked in `instance/mathstream_logs.sqlite`, freshly truncated each run). Files whose score meets or exceeds the threshold are deleted, letting you tune how aggressively to reclaim space. Both staged results and literal caches participate.
+- **Garbage collection** – `collect_garbage(score_threshold)` computes a score from file age, access count, and reference count (tracked in `instance/mathstream_logs.sqlite`, freshly truncated each run). Files whose score meets or exceeds the threshold are deleted, letting you tune how aggressively to reclaim space. Both staged results and literal caches participate. Use `tracked_files()` or `active_streams()` to inspect current state.
 
 Divide-by-zero scenarios raise the custom `DivideByZeroError` so callers can distinguish mathstream issues from Python’s native exceptions.
 
diff --git a/mathstream/__init__.py b/mathstream/__init__.py
index 137182b..8814edf 100644
--- a/mathstream/__init__.py
+++ b/mathstream/__init__.py
@@ -1,11 +1,12 @@
 from .engine import clear_logs, add, sub, mul, div, mod, pow, is_even, is_odd
 from .exceptions import MathStreamError, DivideByZeroError
-from .number import StreamNumber
-from .utils import collect_garbage
+from .number import StreamNumber, free_stream, active_streams
+from .utils import collect_garbage, tracked_files
 
 __all__ = [
     "clear_logs",
     "collect_garbage",
+    "tracked_files",
     "add",
     "sub",
     "mul",
@@ -15,6 +16,8 @@ __all__ = [
     "is_even",
     "is_odd",
     "StreamNumber",
+    "free_stream",
+    "active_streams",
     "MathStreamError",
     "DivideByZeroError",
 ]
diff --git a/mathstream/number.py b/mathstream/number.py
index 4d7a832..1cad0bd 100644
--- a/mathstream/number.py
+++ b/mathstream/number.py
@@ -1,8 +1,15 @@
 import hashlib
+import weakref
+from collections import Counter
 from pathlib import Path
-from typing import Optional, Union
+from typing import Dict, Optional, Union
 
-from .utils import register_log_file, register_reference, touch_log_file
+from .utils import (
+    register_log_file,
+    register_reference,
+    touch_log_file,
+    release_reference,
+)
 
 LOG_DIR = Path("./instance/log")
 
@@ -60,11 +67,19 @@ class StreamNumber:
             if not self.path.exists():
                 raise FileNotFoundError(self.path)
 
+        self.hash = hashlib.sha1(str(self.path).encode()).hexdigest()[:10]
+        self._normalized_path = str(self.path.resolve())
+        self._released = False
+
+        _increment_active(self.path)
+
         if _is_in_log_dir(self.path):
             register_log_file(self.path)
             register_reference(self.path)
 
-        self.hash = hashlib.sha1(str(self.path).encode()).hexdigest()[:10]
+        self._finalizer = weakref.finalize(
+            self, _finalize_instance, self._normalized_path
+        )
 
     def __repr__(self):
         return f"<StreamNumber {self.path.name}>"
@@ -85,3 +100,52 @@ class StreamNumber:
             f.write(data.encode())
         register_log_file(stage_file)
         return stage_file
+
+    def free(self, *, delete_file: bool = True) -> None:
+        """Release this stream's reference and optionally delete the staged file."""
+        if self._released:
+            return
+        self._released = True
+        if self._finalizer.alive:
+            self._finalizer.detach()
+        _decrement_active(Path(self._normalized_path), delete_file=delete_file)
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc, tb):
+        self.free()
+
+
+_ACTIVE_COUNTER: Counter[str] = Counter()
+
+
+def _increment_active(path: Path) -> None:
+    key = str(path.resolve())
+    _ACTIVE_COUNTER[key] += 1
+
+
+def _decrement_active(path: Path, delete_file: bool = True) -> None:
+    key = str(path.resolve())
+    current = _ACTIVE_COUNTER.get(key, 0)
+    if current <= 1:
+        _ACTIVE_COUNTER.pop(key, None)
+    else:
+        _ACTIVE_COUNTER[key] = current - 1
+
+    if _is_in_log_dir(path):
+        release_reference(path, delete_file=delete_file)
+
+
+def _finalize_instance(path_str: str) -> None:
+    _decrement_active(Path(path_str))
+
+
+def free_stream(number: StreamNumber, *, delete_file: bool = True) -> None:
+    """Convenience helper mirroring manual memory management semantics."""
+    number.free(delete_file=delete_file)
+
+
+def active_streams() -> Dict[str, int]:
+    """Return the active StreamNumber paths mapped to in-memory reference counts."""
+    return dict(_ACTIVE_COUNTER)
diff --git a/mathstream/utils.py b/mathstream/utils.py
index af8fdf9..b0d1120 100644
--- a/mathstream/utils.py
+++ b/mathstream/utils.py
@@ -1,7 +1,7 @@
 import sqlite3
 from datetime import datetime, timezone
 from pathlib import Path
-from typing import Iterable, List
+from typing import Iterable, List, Dict
 
 LOG_DB_PATH = Path("./instance/mathstream_logs.sqlite")
 
@@ -178,3 +178,43 @@ def collect_garbage(score_threshold: float) -> list[Path]:
 
     _delete_records(removed)
     return removed
+
+
+def release_reference(path: Path, delete_file: bool = True) -> bool:
+    """Decrease the reference count and optionally delete the file when it hits zero."""
+    normalized = _normalize_paths([path])[0]
+    _ensure_db()
+    with sqlite3.connect(LOG_DB_PATH) as conn:
+        row = conn.execute(
+            "SELECT ref_count FROM refs WHERE path = ?", (normalized,)
+        ).fetchone()
+        if row is None:
+            return False
+        current = row[0] or 0
+        new_count = max(current - 1, 0)
+        if new_count > 0:
+            conn.execute(
+                "UPDATE refs SET ref_count = ? WHERE path = ?", (new_count, normalized)
+            )
+            conn.commit()
+            return False
+        conn.execute("DELETE FROM refs WHERE path = ?", (normalized,))
+        conn.execute("DELETE FROM logs WHERE path = ?", (normalized,))
+        conn.commit()
+
+    removed = False
+    if delete_file and path.exists():
+        try:
+            path.unlink()
+            removed = True
+        except OSError:
+            removed = False
+    return removed
+
+
+def tracked_files() -> Dict[str, int]:
+    """Return a mapping of tracked file paths to their reference counts."""
+    _ensure_db()
+    with sqlite3.connect(LOG_DB_PATH) as conn:
+        rows = conn.execute("SELECT path, ref_count FROM refs").fetchall()
+    return {path: ref_count for path, ref_count in rows}