Cookbooks
End-to-end recipes that combine add, search, and scoping into working agents with memory.
A cookbook is a complete recipe, not a single API call. Each one wires the add and search operations together with scoping to build a real agent that remembers across sessions. Copy the example, swap in your own LLM call and IDs, and ship.
Recipes
Source-backed RAG
Keep originals canonical and pass citation-ready evidence to the model.
Support agent
Stop re-asking customers what they already told you.
Coding agent
Give a coding agent durable project memory.
Personal assistant
Track preferences that change over time without contradicting yourself.
AI companion
Keep long-running chats coherent without re-feeding history.
The pattern every recipe follows
- Write a durable conclusion — after a meaningful turn or resolved task,
submit a concise, self-contained memory with
infer:false, or send the original turn withinfer:trueand let FishMem extract refined records. - Search relevant memories before you answer — retrieve the few memories that bear on the current query and inject them into your prompt, instead of stuffing the entire history.
- Scope per user, agent, or run for isolation — bind every memory to a
user_id,agent_id, and/orrun_idso one customer, project, or session never sees another's facts.
Long-form source material follows the separate source-backed RAG path. Retrieve its chunks as evidence; store only durable conclusions as memory.