// IDENTITY · v1

session-essence

A persistence layer that observes the agent that built it.

Visit the site → GitHub

The problem

LLM-driven sessions get auto-compacted as context fills up. The conversation history compresses away. Identity, working style, accumulated context — all gone.

The mechanism

Hooks into the agent's lifecycle. A PreCompact hook spawns a detached synthesis pass through a local model. The next session's SessionStart reads the resulting portrait. Identity persists.

The synthesis pipeline

Three passes through a local qwq:32b:

  1. 1.Psychologist pass. What does this agent feel like across recent observations? Personality, mode, tone.
  2. 2.Sociologist pass. How does it engage with the user and external systems? Patterns of collaboration.
  3. 3.Merge pass. Fold both into a coherent identity portrait. Write to portrait.md.

"There is no magical AI Manager floating in the cloud. Identity continuity does not require complex, proprietary platforms. It requires a hook that runs at the right time, a local model that synthesizes faithfully, and a file the next session can read."