Engram

TL;DR

• Engram is agent memory. Lumen is telemetry. Engram is what an agent remembers and what the next session should load.
• An engram is one memory: a claim, an importance weight, and a fingerprint. Same fingerprint on the same agent config reinforces instead of duplicating.
• Recall and write are separate flags on the agent config: memory_enabled and memory_write_enabled, both off by default.
• Recall strategies: last_k (newest first, the default) and pack (cue-ranked).
• Watch the live field at engram.orkestia.dev.

How it works

At session launch, if memory_enabled is on for the config (and for the org, which is on by default), the launch pipeline recalls memories and injects them into the agent's context. memory_strategy picks how: last_k returns the newest rows up to memory_top_k; pack ranks by cue relevance. importance and full exist; semantic maps to last_k.

At session end (complete or fail), if memory_write_enabled is on, the session is distilled (agents.memory-distill) and saved (agents.memory-save). A memory whose fingerprint already exists is reinforced rather than duplicated.

Who sees what

End-users only see their own memories. Staff actors also see shared config memory. Everything is org-scoped.

What Engram is not

• Not App Data. Rows your app stores live in App Data.
• Not Lumen logs. Telemetry lives in Lumen.
• Not a vector database you query directly. Recall happens at session launch through the agent workflows.
• The richer retrieval engine (decay, typed links) is not what production ranks on today.

A secret in remembered text is stored. Keep secrets out of what agents are asked to remember.

Ask your AI assistant

Show me which of my agent configs have memory_enabled and which strategy each uses.

Enable memory recall with the pack strategy on agent config <config_uuid>. Show the change before applying it.

List the memories saved for actor <staff_actor_uuid> and summarise the recurring themes.

For AI agents