Skip to content

Build a project's memory

Memory is an active layer: decisions and learnings produced by your sessions are captured, consolidated and reinjected into upcoming sessions. The agent does not start from scratch.


The memory-local addon listens to chat events:

  • Session end (chat.session.ended) — the addon analyzes the transcript and captures milestones (decisions, learnings, actions, methods).
  • Runtime failure — notable errors are logged.
  • Explicit command (e.g. /memorize) — you request a specific capture.

Entries are sorted by importance level:

LevelTypical content
L1Short note, observation
L2Relevant technical detail
L3Session synthesis
L4Decision, conclusion
L5Structured learning, method

Memory view in the project:

  • List of entries sorted by level and date.
  • Filtering by tag, source session, agent.
  • Detail of each entry (summary, content, actions, decisions).

Before each turn, memory-local enriches the agent’s context with:

  • Recent entries relevant to the project.
  • Still-valid decisions and learnings.
  • A condensed recall pack that fits the available token budget.

You see the effect immediately: the agent recalls what was decided without you having to re-paste.


  • No secrets, no credentials. If a secret appears in a session, it must not be captured. The addon applies conservative filters.
  • Not the raw history. Memory is a semantic extraction, not a transcription log.

The Addons view lets you disable memory-local per project if you don’t want capture on this project. Sessions run normally, simply no entry will be added.