Memory & Session Search

Kyber has two complementary memory systems for persistence and recall.

Persistent Memory

The memory tool manages curated, long-term memory stored in markdown files:

• MEMORY.md — The agent’s personal notes, observations, and learnings
• USER.md — What the agent knows about you (preferences, habits, context)

How to use

Just ask the agent to remember something:

Remember that I prefer TypeScript over JavaScript

Save this: my project uses PostgreSQL not MySQL

The agent will use the memory tool to add entries to the appropriate file.

Memory limits

Memory files are bounded to prevent context bloat:

• MEMORY.md: ~2,200 characters
• USER.md: ~1,375 characters

When limits are reached, older or less important entries should be consolidated. The agent handles this automatically.

Session Search

The session_search tool searches through past conversations. This is useful when:

• You reference something from a previous chat (“Remember when we fixed that docker issue?”)
• You want to recall how something was done before
• You’re troubleshooting and want to see what worked previously

How to use

Search our past convos for anything about nginx

What did we do about the database migration issue?

The agent will use session_search to find relevant past conversations.

How it works

Sessions are stored as JSONL files in ~/.kyber/sessions/. The search tool:

1. Scans all session files for matching keywords
2. Returns the most relevant conversations with context
3. Shows a summary of what was discussed

This gives the agent a “long-term memory” of your interactions without bloating the current context.