Memory overview · HALO Docs

HALO Memory turns end-user conversations into project-isolated topics, summaries, and linked raw exchanges. Retrieval returns relevant source records to your model; HALO does not generate the final user-facing answer.

#Scope model

Value	Meaning	Rule
`HALO_API_KEY`	Authenticates the trusted caller	Header only; never a tool argument
`projectKey`	Memory product or environment	Must reference an existing project and must not begin with sk-
`endUserKey`	Your stable customer identifier	Required and non-null
`sessionData`	Current messages and application state	Pass when the current task needs context selection

#Recommended flow

1
Declare the retrieve function
Add halo_retrieve_end_user_memory to the tools sent to your model.
2
Execute only when called
Your backend inserts project and end-user identity, then calls the HALO function endpoint.
3
Return sources to the model
Feed functionResponseback into the model's tool loop.
4
Capture the final exchange
After the assistant answer is complete, write the user/assistant exchange separately.

#What Memory stores

An overall summary for an end-user scope.
Topic keys, labels, aliases, confidence, and summaries.
Raw user/assistant request and response records linked to topics.
Processing jobs and usage needed to classify and summarize writes.

#Integration modes

Function calling

Best for apps that keep their existing LLM client and tool loop.

Direct REST

Best for dashboards, maintenance jobs, and explicit data flows.

Router headers

Legacy model-proxy mode can request capture or injected retrieval.