maximem-vity-sdk

Vity by Maximem AI — a lightweight Python SDK for cross-session semantic memory. Store facts and preferences, recall relevant context for a prompt, run semantic search, and capture whole conversation turns — all over the Maximem REST API.

from maximem_vity import VityClient

client = VityClient(api_key="mx_...")

# Recall context ready to inject into a system prompt
context = client.recall(current_prompt="What color theme does the user like?")

# Store an explicit fact
client.store("User prefers dark mode", memory_type="preference")

# Semantic search
hits = client.search("editor preferences", top_k=5)

Installation

pip install maximem-vity-sdk

Requires Python ≥ 3.9. The only runtime dependency is httpx.

Authentication

Every request is authenticated with a Maximem API key (it starts with mx_). The API key owns the memory space — VityClient takes no channel or user ID. Use a separate API key whenever memories must be isolated between users or accounts.

Get your key at maximem.ai/dashboard.

import os
from maximem_vity import VityClient

client = VityClient(api_key=os.environ["MAXIMEM_API_KEY"])

export MAXIMEM_API_KEY="mx_..."

API

recall(current_prompt, recent_messages=None, max_tokens=1000, strategy="hybrid") -> str

Returns a plain-text context block for the current prompt, ready to inject into a system prompt (empty string if nothing relevant). strategy is one of "semantic", "recency", or "hybrid".

context = client.recall(
    current_prompt="Remind me what we decided about the API.",
    recent_messages=[{"role": "user", "content": "Let's revisit the API."}],
    max_tokens=800,
)

store(content, memory_type="fact", category="", importance="medium") -> dict

Stores a single explicit memory. memory_type accepts the high-level aliases fact, preference, emotion, episode, knowledge, profile (mapped to Maximem categories automatically), or pass category directly to override. importance is "low", "medium", or "high". Returns {"id": ..., "stored": True}.

client.store("User's timezone is IST", memory_type="fact", importance="high")

search(query, top_k=10, limit=0, category="", min_score=0.0) -> list[dict]

Semantic search. Returns a list of dicts shaped as {"content", "type", "score", "id", "created_at"}. top_k is the convenience limit (capped at 20 by the API); limit overrides it when set.

for hit in client.search("favourite languages", top_k=5, min_score=0.3):
    print(hit["score"], hit["content"])

capture(messages, agent_id="hermes", session_key="") -> dict

Captures a full conversation turn into long-term memory. messages is a list of {"role", "content", "timestamp"} dicts (timestamps are filled in if omitted). Returns {"captured": N, "deduplicated": M}.

client.capture([
    {"role": "user", "content": "I hate comic sans"},
    {"role": "assistant", "content": "Noted — I'll avoid it."},
])

forget(query="", dry_run=True) -> dict

Deletes memories matching a query. Defaults to a dry run — pass dry_run=False to actually delete. Returns {"count": N, "ids": [...]}.

preview = client.forget(query="old project notes")          # safe preview
client.forget(query="old project notes", dry_run=False)     # actually delete

Convenience shims

• ingest(messages, session_id="") — alias for capture().
• get_profile() — a broad search returning the user's stored memories.

Lifecycle

VityClient holds a pooled HTTP connection. Close it explicitly or use it as a context manager:

with VityClient(api_key="mx_...") as client:
    client.store("User prefers tabs over spaces")
# connection closed automatically

Error handling

All errors derive from VityError:

from maximem_vity import VityClient, VityError

try:
    client.store("...")
except VityError as e:
    logger.warning("Memory write failed: %s", e)

Configuration

Development

pip install -e ".[dev]"
pytest                      # run the test suite
python -m build             # build the wheel + sdist into dist/

License

MIT © Maximem AI