Memory

Memory is long-term, semantic recall across sessions — distinct from a checkpoint store (durable serialization of one run). It's a 3-method protocol: save / search / load.

Pick a memory type

Type	Recall	Install
InMemoryVectorMemory	embedding cosine similarity	spinekit
BufferMemory	recency (last-N), non-semantic	spinekit
PgVectorMemory	Postgres + pgvector, at scale	spinekit[postgres]

from spine_backends import InMemoryVectorMemory, BufferMemory

mem = InMemoryVectorMemory()        # semantic
mem = BufferMemory()                # simple recency

Choose how text is embedded

Memory takes any Embedder — async def embed(text) -> list[float]. Swap the default offline embedder for a real model, or bring your own:

from spine_backends import InMemoryVectorMemory, HashEmbedder, OpenAIEmbedder

# offline default — deterministic, no API
mem = InMemoryVectorMemory(embedder=HashEmbedder())

# real OpenAI embeddings
mem = InMemoryVectorMemory(embedder=OpenAIEmbedder("text-embedding-3-small"), dim=1536)

# your own (sentence-transformers, Cohere, local …)
class MyEmbedder:
    async def embed(self, text: str) -> list[float]:
        return my_model.encode(text).tolist()

mem = InMemoryVectorMemory(embedder=MyEmbedder())

Use it directly

await mem.save("The launch code is alpha-seven.", session_id="s1")
hits = await mem.search("what is the launch code?", k=3)
print(hits[0].record.content, hits[0].score)

Wire it into an agent

MemoryRecall searches memory by the user's question, injects matches as an ephemeral system message (non-destructive), and saves the exchange afterward:

from spine_core import Agent
from spine_middleware import MemoryRecall

agent = Agent("openai:gpt-4o-mini", middleware=[MemoryRecall(mem, k=3)])

Scope recall to the current session with MemoryRecall(mem, scope_session=True).

pgvector at scale

from spine_backends import PgVectorMemory, OpenAIEmbedder

mem = PgVectorMemory(
    "postgresql://localhost/app",
    embedder=OpenAIEmbedder(),
    dim=1536,
)

Requires the pgvector extension in the database; the table is created on first use.