Memory

Neural Memory Operations

Operations

• search <query> - Search lessons
• store <lesson> - Store a lesson manually
• stats - Memory statistics
• consolidate - Force consolidation
• decay - Force decay/pruning

When the user asks for a memory operation, infer the operation and arguments from the prompt. If the operation is ambiguous, ask one concise clarifying question before calling memory MCP tools.

Subcommands

search

Search both Graphiti (episodic) and Qdrant (semantic):

1. Call mcp__graphiti__search_nodes with the query
2. Call mcp__qdrant__qdrant-find with the query
3. Merge and deduplicate results
4. Present ranked by relevance

store

Manually store a lesson:

1. Parse the lesson text from arguments
2. Call mcp__graphiti__add_episode with the lesson
3. Call mcp__qdrant__qdrant-store with the lesson
4. Confirm storage with node/vector IDs

stats

Display memory statistics:

1. Call mcp__graphiti__get_status for graph stats
2. Report: node count, edge count, episode count
3. Show last consolidation timestamp
4. Show log file sizes from ~/.codex/fusengine/logs/00-memory/

consolidate

Force episodic to semantic consolidation:

1. Read recent episodes from Graphiti
2. Identify patterns (recurring errors, common solutions)
3. Store consolidated patterns in Qdrant
4. Report consolidation summary

decay

Force memory decay and pruning:

1. Apply power law decay: R(t) = strength * (1+t)^(-rate)
2. Prune nodes with retention < 0.10
3. Prune edges with weight < 0.01
4. Report pruned count