Compact

Multi-level summary system for enhanced search

Compact generates AI summaries from your conversations to improve search quality. Inspired by claude-mem.

Optional feature — everything works without it. Your original data is never modified; Compact adds summaries on top.

Level System

Memex organizes conversation data into levels:

Level	Name	Granularity	Description
L0	Raw messages	Every message	Original content, always preserved
L1	Observations	Per tool call	What tools were used and results
L2	Talk summaries	Per prompt-response	Summary of each conversation turn
L3	Session summaries	Per session	Overview of entire session

L0 is always available — this is what "Never Compacted" protects. L1/L2/L3 are generated by Compact.

How It Works

Compact uses two Ollama models:

Chat model (default: qwen3:0.6b) — generates L1/L2/L3 summaries from raw messages
Embedding model (default: bge-m3) — converts summaries to vectors for semantic search

L0 (raw messages)
 │
 ├──▶ L1 (observations) ──▶ L2 (talk summaries) ──▶ L3 (session summaries)
 │                                │                        │
 └────────────────────────────────┴────────────────────────┘
                                  │
                                  ▼
                          Embedding model
                                  │
                                  ▼
                          LanceDB (vectors)

Generation flow:

L1 is generated from L0 tool calls
L2 is generated from L0 messages + L1 observations
L3 is generated from L2 summaries

L3 depends on L2. To get session summaries, both l2_talk_summary and l3_session_summary must be enabled.

How to Search

MCP search_history uses the level parameter to choose what to search:

Value	Searches	Best for
`sessions` (default)	L3 summaries	Quick overview, finding relevant sessions
`talks`	L2 summaries	Implementation details, specific solutions
`raw`	L0 messages	Exact wording, precise recall

Fallback: If no results at current level, automatically tries the next: sessions → talks → raw.

Why no L1 in MCP? L1 (tool-level observations) is available for context injection, but not exposed in MCP search — it's more useful for automatic context retrieval than manual search.

Context Injection

Compact also powers the Hooks feature for automatic context injection:

Hook	Uses	What's Injected
SessionStart	L3 summaries	Recent session overviews
UserPromptSubmit	L2 summaries + vectors	Relevant conversation context

Without Compact, SessionStart has limited context. UserPromptSubmit requires both Compact and embedding model for vector search.

Enabling Compact

Compact is disabled by default. See Configuration to enable.