How it works
Every memory is automatically classified into a category when it is stored. Instead of calling an LLM to categorize, MemWire uses cosine similarity to anchor embeddings:
- Each category has one or more example sentences (anchors) that define its meaning
- At startup, anchor embeddings are computed as the centroid of all examples for that category
- When a new memory arrives, its embedding is compared to every anchor centroid
- The category with the highest cosine similarity wins — as long as it clears
classification_threshold
This is fully local, runs in microseconds, and requires no API key.
Default categories
| Category | Example anchors |
|---|
fact | ”This is a factual statement”, “The capital of France is Paris” |
preference | ”I prefer dark mode over light mode”, “I like my coffee black” |
instruction | ”Always respond in formal English”, “Never share my personal information” |
event | ”We had a team meeting yesterday”, “The product launched last week” |
entity | ”This is about a company called Acme”, “John is the head of engineering” |
Code example
from memwire import MemWire, MemWireConfig
config = MemWireConfig(qdrant_path="./memwire_data")
memory = MemWire(config=config)
records = memory.add(
user_id="alice",
messages=[
{"role": "user", "content": "I always prefer bullet points over long paragraphs."},
{"role": "user", "content": "Our Q1 planning session is scheduled for next Monday."},
{"role": "user", "content": "Water freezes at 0 degrees Celsius."},
],
)
for record in records:
print(f"{record.category}: {record.content}")
# preference: I always prefer bullet points...
# event: Our Q1 planning session...
# fact: Water freezes at...
Filtering search by category
results = memory.search(
"formatting preferences",
user_id="alice",
category="preference",
top_k=5,
)
for record, score in results:
print(f"[{score:.2f}] {record.content}")
Adding custom categories
Define domain-specific categories for your application:
memory.add_anchor(
name="bug_report",
text="There is a crash or error in the application",
user_id="alice",
)
memory.add_anchor(
name="feature_request",
text="The user is asking for a new capability to be added",
user_id="alice",
)
add_anchor with the same name accumulate examples — the centroid is recomputed each time, sharpening the category boundary.
Customising anchors at config level
config = MemWireConfig(
qdrant_path="./memwire_data",
classification_threshold=0.05, # lower = more memories get classified
default_anchors={
"goal": [
"This is something the user wants to achieve",
"I want to finish the project by end of quarter",
],
"constraint": [
"This is a limitation or restriction",
"We only have a budget of $10,000",
],
}
)
Setting default_anchors replaces the built-in set entirely. Include any categories you want to keep from the defaults.
Configuration reference
| Parameter | Default | Description |
|---|
classification_threshold | 0.05 | Minimum cosine similarity to assign a category. Below this the memory is tagged unknown. |
default_anchors | 5 built-in categories | Dict of category_name → list[example_sentences]. |
