Memory

Memory lets the agent store and retrieve facts across steps — or across runs.

step N: memory_add("user prefers dark mode")
step N+5: memory_search("user preferences")
           → ["user prefers dark mode"]  score=0.94

The problem it solves

Each loop iteration the agent only sees its task description and tool results from the current session. Long tasks that need to recall earlier findings — or agents that run repeatedly and need continuity — require explicit memory.

Two backends

Backend	Persists?	Setup	Use when
InMemoryStore	No (resets on exit)	Zero	Within a single run
ChromaMemory	Yes (disk)	pip install 'gantrygraph[memory]'	Across multiple runs

InMemoryStore

from gantrygraph import GantryEngine
from gantrygraph.memory import InMemoryStore

agent = GantryEngine(
    llm=...,
    memory=InMemoryStore(),
    max_steps=30,
)

ChromaMemory

from gantrygraph.memory import ChromaMemory

agent = GantryEngine(
    llm=...,
    memory=ChromaMemory(
        collection_name="my_agent",
        persist_directory="/var/lib/agent/memory",
    ),
)

How it works

When you attach a memory store, the engine adds two tools to the agent's tool list:

Tool	What it does
memory_add	Store a text snippet with optional metadata
memory_search	Return the k most semantically similar entries

The agent calls these tools on its own — no extra instructions needed. Search uses vector similarity (cosine or L2 depending on backend).

Similarity score

Results are returned with a score between 0 and 1. Higher = more relevant.

results = await memory.search("authentication errors", k=3)
for r in results:
    print(f"{r.score:.2f}  {r.text}")

Custom backend

Subclass BaseMemory:

from gantrygraph.memory.base import BaseMemory, MemoryResult

class RedisMemory(BaseMemory):
    async def add(self, text: str, metadata=None) -> None: ...
    async def search(self, query: str, k: int = 5) -> list[MemoryResult]: ...
    async def close(self) -> None: ...

---

See also: Memory guide · API reference