Manage what Jacky remembers about your past conversations and environment context. Jacky utilizes two distinct types of memory: Short-term Conversation Memory (a rolling window of the most recent turns) and Long-term Knowledge (persistent facts extracted over time).
Conversation Memory
Jacky keeps a rolling history of recent user-assistant interactions and injects them into every chat request as proper multi-turn messages. This lets you have natural, context-aware follow-up conversations.
To save context length and prevent LLM confusion, conversation memory is scoped to direct questions only. Idle comments, auto-reactions (like battery warnings, music feedback, or collectibles), and screen-interaction systems are automatically excluded.
Configuration Options (memory)
| Option in UI | Config Key | Range / Type | Description |
|---|---|---|---|
| Turns in memory | memory_max_turns | 1 – 50 (Default: 5) | Number of user ↔ assistant exchange pairs to keep in the rolling context. |
| Save memory to disk | memory_persist | Boolean (Default: false) | If checked, history is saved to memory.json on disk and survives restarts. |
Long-term Knowledge
Beyond the rolling conversation window, Jacky can extract and retain persistent facts about you across sessions—such as your name, preferences, job, projects, or tools. These facts are formatted as a compact list and injected into every LLM request.
Dual-Gated Privacy: This feature is dual-gated for privacy. It only runs when Enabled is checked here and the Allow knowledge memory permission is enabled in the Permissions tab.
Configuration Options
| Option in UI | Config Key | Range / Type | Description |
|---|---|---|---|
| Enabled | knowledge_enabled | Boolean (Default: true) | Master toggle for long-term knowledge extraction and injection. |
| Extract every N user turns | knowledge_extract_every | 1 – 50 (Default: 5) | How frequently a background LLM process runs to extract new facts from logs. |
| Max. facts | knowledge_max_facts | 5 – 200 (Default: 50) | Maximum stored facts before triggering a consolidation pass. |
| Token budget | knowledge_token_budget | 50 – 2000 (Default: 500) | Approx. token budget (≈4 chars/token) for the injected knowledge block. |
Manual Operations & Editing
- Facts List: An editable text area containing the list of learned facts (stored in
knowledge.json).- Edit the list directly—one fact per line.
- Optionally prefix facts with a category in brackets, like
[identity],[work], or[preferences](e.g.,[preferences] The user prefers the color red).
- Extract now: Manually triggers the background LLM extraction on your recent conversation logs immediately.
- Forget all: Instantly wipes the local
knowledge.jsonfile, deleting all stored facts.