Personal assistants
Preferences that survive a new chat thread. Dietary rules, tone, and calendar bounds return as a packet the model can cite.
Product
A memory service with an export path.
Add memories from conversation. Search by meaning and by entity. Patch when the world changes. Export when the user asks to leave.
HTTP agents can call Neemel. You can keep your current orchestrator.
Cloud
Hosted add and search
Usage-based billing on writes and recalls. Isolation per project. Optional EU region. Webhooks fire when a fact is superseded so your agent can refresh the packet on the next turn.
US East (Ashburn) and EU West (Frankfurt) keep memory payloads in-region. Estate contracts can pin a single region and a private endpoint.
Two production regions. Trace uses US East. Keep can enable both.
Graph
Who, what, and when
Hybrid store: graph for who, what, and when. Vectors for fuzzy recall. Key-value for fast flags. Temporal edges so last week’s plan reads as history.
People, preferences, and time arrows live as typed nodes. A faded diary line becomes a small structured card the agent can cite.
user mem_0f2a
Mira Kade is allergic to shellfish. Confirmed 3 March 2026. Source: onboarding chat usr_mira.
Open source
SDK and local store
SDK and local store for development. Cloud is the production path when you want the graph operated for you. The contribution guide lives in Docs · Contributing.
Public SDK under Apache-2.0. 890,000 pip installs in 2026. Contribution guide.
pip install neemel from neemel import Client client = Client(api_key="nm_live_7c91e4a0b2f65d18c3a9e80d") client.memories.add(user_id="usr_mira", text="Prefers concise replies.")
Use cases
Four desks that keep a mnemonic
Support agents
Account facts retrieved as a packet: plan, last ticket, promised callback. The spoken or typed turn stays short.
Multi-agent crews
A shared mnemonic with scoped writes. Planner, researcher, and reviewer read the same graph and write under agent tags.
Voice agents
Low-latency retrieve so the spoken turn stays short. p95 search in US East is 38ms on Keep.
Customers
Built for developers who already have an agent
Neemel is for teams that need an agent to persist. Consumer chat apps, B2B copilots, and internal ops agents. Role titles until legal clears logos.
Agent platform engineer
Northline cut prompt size on a consumer assistant by retrieving a 12-line packet.
Applied ML
Harbor Ops marked travel plans as retired after the trip ended. The next session stopped offering last week’s itinerary.
Voice runtime lead
Kite Desk holds a 900ms spoken budget. Retrieve stays inside 40ms so TTS still owns the rest of the turn.
Integrations
Harness types
Logo-free list of harness types Neemel speaks today.
HTTP agents
REST from any runtime. Bearer key, JSON body, ranked memories back.
Tool-calling loops
Expose add, search, and patch as tools. The model decides when a fact is durable.
Voice runtimes
Search on barge-in. Keep the packet under 800 tokens for a spoken reply.
MCP clients
Neemel ships an MCP server so a desktop or IDE agent can read and write memory with scoped tools.