Three patterns teams reach for first. Each uses the same five-method API:
add, search, get, get_all,
delete.
Before each turn, search memories for the customer’s plan tier, past issues, and tone preferences—inject the top facts into the system prompt.
hits = memory.search(
message,
filters={"user_id": customer_id},
top_k=5,
)
context = "\n".join(h["memory"] for h in hits["results"])
After responding, append the exchange with memory.add so the next login
still knows location, stack choices, and standing instructions.
memory.add(
[{"role": "user", "content": msg},
{"role": "assistant", "content": reply}],
user_id=user_id,
)
Scope episodic tool outcomes with run_id while keeping stable user
preferences under user_id for cross-run recall.
memory.add(
tool_summary,
user_id="team",
run_id=run_id,
infer=False,
)Browser demo • zero backend • localStorage
Add facts, scope them, search with hybrid ranking, and run a full retrieve → respond → remember loop. The retrieval and scoring mirror Recollect's local embedder + hybrid search (semantic + keyword + entity).
This demo replicates the local-dev path (RecollectConfig.local_dev()) entirely in the browser using the same hybrid scoring approach.
The same flow as examples/chat_with_memory.py: retrieve, answer, then write
back new facts. The live demo above executes an equivalent loop using the real retrieval logic.
from openai import OpenAI
from recollect import Memory
memory = Memory()
client = OpenAI()
user_id = "demo_user"
def reply(message: str) -> str:
retrieved = memory.search(message, filters={"user_id": user_id}, top_k=5)
memories = "\n".join(f"- {r['memory']}" for r in retrieved["results"])
system = f"You are helpful. User memories:\n{memories or '- (none)'}"
answer = client.chat.completions.create(
model="gpt-4o-mini",
messages=[
{"role": "system", "content": system},
{"role": "user", "content": message},
],
).choices[0].message.content
memory.add(
[{"role": "user", "content": message},
{"role": "assistant", "content": answer}],
user_id=user_id,
)
return answerNo API key required for search/add smoke tests—use the local embedder and infer=False.