Memory

Persistent memory: session search, curation, summarization, and user modeling.

Synapse does not rely on the context window alone. It maintains memory that survives sessions through three mechanisms, each inspectable and editable — nothing about your memory is hidden from you.

Conversation history is stored with SQLite FTS5 full-text search. Relevant past sessions are retrieved on demand and condensed with LLM summarization, so recalled context stays compact instead of dumping raw transcripts.

How recall flows

StageWhat happens
ConversationEvery session produces text worth keeping
Session storeHistory persisted with full-text search (FTS5)
SearchRelevant past sessions retrieved on demand
Relevant memoryLLM summarization condenses what matters
Current taskContext is applied where it helps

Agent-curated memories

The agent curates its own memory files (MEMORY.md / USER.md style entries) and periodically nudges itself to persist important knowledge. You can read and edit these files directly — they live in your Synapse home, in plain Markdown.

User modeling

Optional Honcho integration builds a dialectic model of you across sessions, deepening personalization over time: how you like answers structured, what projects you are working on, which conventions you follow.

  • /usage — check current context consumption
  • /compress — compact the active context
  • /insights [--days N] — review recent session activity
  • /memory — inspect curated memory state