01 · Session context
Working memory
Rolling events with configurable TTL and per-user caps.- Ephemeral session state
- Automatic pruning
Memory infrastructure for AI
Memory that moves at thought speed.
A Rust-backed memory engine for Python AI assistants. Get 1–10 ms in-RAM recall, durable SQL persistence, and zero required services.
- Rust + Python
- SQLite by default
- MIT Licensed
Working memory
Ephemeral · sub-second
Short-term
Recent · seconds → hours
Long-term
Durable · days → years
Unified memory core
Neural recall · 1–10 ms
One system. Three memory horizons.
Fast where it matters. Durable where it counts.
Every query stays on the Rust-backed in-memory index. SQL handles durability quietly in the background, while short-term and working memory keep immediate context close.
02 · Recent context
Short-term memory
TTL-managed traces for what matters right now.- 15-minute default TTL
- Optional by configuration
03 · Durable recall
Long-term memory
Persistent traces and embeddings that survive restarts.- SQLite or PostgreSQL
- Hydrates directly to RAM
RustMemoryIndex
Recall in 1–10 ms
Cosine similarity over an in-RAM vector index. No database round-trip on the critical path.- Query path
- RAM only
- Persistence
- One SQL database
- Runtime
- Async-safe
Built for the hot path
Database durability without database latency.
MemoryCore splits persistence from retrieval. Writes update SQL and the index together. Reads stay in RAM. Store I/O runs in a worker thread, so your async event loop stays responsive.
1–10 ms
Typical in-memory recall
0
Required external services
64-bit
Structured memory key
3
Integrated memory layers
Live request trace
index healthy
-
0.0 ms
receivequery + user_id✓
-
0.7 ms
embedvector ready✓
-
4.2 ms
recallRust vector index5 hits
-
6.8 ms
returnranked memories✓
Pythonic by design
From zero to durable memory in a few lines.
SQLite ships with Python and the schema creates itself. Bring an embedder—or pass vectors directly—and start remembering.
from memory_core import build_memory_system
system = build_memory_system()
await system.remember(
user_id="alice",
summary="Prefers Python for scripting tasks",
importance=0.8,
tags=["preferences", "programming"],
embedding=my_vector,
)
result = await system.recall(
user_id="alice",
query_text="programming languages",
query_embedding=query_vector,
limit=5,
)
persisted to SQLite
Rust index updated
ready to recall
Bring your embedder
Implement one async protocol and use raw text everywhere. MemoryCore handles vector generation at the boundary.
Filter before similarity
The optional SMK assistant index packs topic, kind, tools, difficulty, generality, and importance into a 64-bit key.
Swap persistence later
Start locally with SQLite. Point one URL at PostgreSQL when the application grows—without changing the memory API.
Give your assistant continuity
Memory should feel instant.
Now it can.
Open source, MIT licensed, and designed for the Python applications you already build.