Write path
A conversation lands as turns. Neemel keeps the lines that should outlive the thread: a preference, an account fact, a plan with a date. Retired facts get a temporal edge so last week’s plan reads as history.
Memory layer for agents
Neemel stores, updates, and retrieves what should persist across sessions: preferences, facts, and facts that have been retired. Your model receives a current mnemonic, sized for the turn.
Loop
Ingest messages. Extract durable facts. Write to graph and vectors. Retrieve a small, current packet at inference time.
A conversation lands as turns. Neemel keeps the lines that should outlive the thread: a preference, an account fact, a plan with a date. Retired facts get a temporal edge so last week’s plan reads as history.
At inference, search returns ranked memories plus an optional context block. The packet is scoped to user, session, or agent. HTTP agents can call Neemel. You can keep your current orchestrator.
Three steps
POST a message, a tool result, or a typed fact. Scopes mark who may read it later.
Extraction writes entities, relations, and embeddings. A superseded line is patched in place.
Search by meaning and by entity. The model sees a current mnemonic, sized for the turn.
Product
Metered add, search, and update. Isolation per project. Optional EU region. Webhooks when a fact is superseded.
Run the framework in your VPC. SDK and local store for development. Cloud is the production path when you want the graph operated for you.
Entities, relations, time. Hybrid store: graph for who, what, and when. Vectors for fuzzy recall. Key-value for fast flags.
Where it lands
Measured
Notes
Notes on retrieval quality, latency, and how superseded facts are marked. Read the method notes.
“We cut prompt size. The agent still knew the user’s constraints.”Leah Ortiz, agent platform engineer, Northline
“Retired facts stopped haunting the next session.”Samir Adeyemi, applied ML, Harbor Ops
Install the SDK, add a fact, retrieve a packet. The first 10,000 memories live on Trace.